Brushless controller data analysis method and system

Abstract

The invention relates to the technical field of brushless controllers, and discloses a brushless controller data analysis method and system.The brushless controller data analysis method comprises the steps that firstly, high-frequency current data and low-frequency temperature data are fused to construct a brushless controller state multi-mode characteristic spectrum; then automatically extracting a brushless controller state multi-modal coding deep feature map from the multi-modal map by using a convolutional neural network; and inputting the brushless controller state multi-mode coding deep feature map into a parallel fault attention branch to obtain a set of fault exclusive feature vectors, thereby successfully decomposing a complex composite fault problem into a plurality of parallel and independent single fault identification tasks, and finally performing independent probability prediction based on respective exclusive features. Therefore, multi-label composite fault diagnosis is realized accurately and reliably. In this way, precise decoupling and diagnosis of the composite fault of the brushless controller are achieved, and therefore the operation reliability and safety of a power system in key equipment are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of brushless controller technology, and more specifically, to a brushless controller data analysis method and system. Background Technology

[0002] Brushless DC motors and their controllers, with their advantages of high efficiency, high power density, and long lifespan, have been widely used in industrial automation, new energy vehicles, aerospace, robotics, and high-end consumer electronics. As applications increasingly demand higher system reliability, ensuring the stable operation of controllers under complex conditions has become a key technological challenge. Failure to diagnose any potential fault in a timely manner can lead to equipment downtime or even serious safety accidents; therefore, achieving efficient and accurate fault diagnosis is crucial.

[0003] Existing diagnostic methods mainly include physical models based on threshold rules and machine learning techniques based on signal features. Early methods relied on preset current, voltage, or temperature thresholds for judgment, which were logically simple but lacked adaptability, making it difficult to identify early faults and prone to false alarms or missed alarms. In recent years, diagnostic methods combining Fourier transform and wavelet transform to extract signal features and utilizing classifiers such as support vector machines and decision trees have improved accuracy. However, these methods rely on manual feature engineering, require extensive domain knowledge, and have limited feature representation capabilities when facing complex faults with varying operating conditions or multiple concurrent faults, resulting in decreased diagnostic effectiveness. Therefore, existing technologies still have significant shortcomings in automation, generalization capabilities, and complex fault identification, and there is an urgent need for more intelligent and robust diagnostic solutions. Summary of the Invention

[0004] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a brushless controller data analysis method and system.

[0005] According to one aspect of this application, a brushless controller data analysis method is provided, comprising: acquiring high-frequency current data and low-frequency temperature data; performing feature extraction and feature merging on the high-frequency current data and low-frequency temperature data to obtain a brushless controller state multimodal feature map; inputting the brushless controller state multimodal feature map into a backbone network based on a convolutional neural network model to obtain a brushless controller state multimodal encoding deep feature map; inputting the brushless controller state multimodal encoding deep feature map into a parallel fault attention branch to obtain a set of fault-specific feature vectors; performing independent fault probability prediction on each fault-specific feature vector in the set of fault-specific feature vectors to obtain a fault probability vector; and generating multi-label diagnostic results based on the fault probability vectors to obtain a composite fault diagnosis result.

[0006] In one possible implementation, feature extraction and feature merging are performed on high-frequency current data and low-frequency temperature data to obtain a brushless controller state multimodal feature map, including: extracting a current analysis window from the high-frequency current data; performing a short-time Fourier transform on the current analysis window to obtain a current time-frequency spectrum; calculating the mean of the low-frequency temperature data to obtain aggregated temperature values; creating a two-dimensional matrix with the same dimension as the current time-frequency spectrum, and filling each position of the two-dimensional matrix with the aggregated temperature values ​​to obtain a temperature distribution map; merging the current time-frequency spectrum and the temperature distribution map along the channel dimension to obtain the brushless controller state multimodal feature map.

[0007] In one possible implementation, the backbone network of the convolutional neural network model is the MobileNet model.

[0008] In one possible implementation, inputting the deep feature map of the brushless controller state multimodal encoding into a parallel fault attention branch to obtain a set of fault-specific feature vectors includes: inputting the deep feature map of the brushless controller state multimodal encoding into a first fault attention branch to obtain a first attention weight map; calculating a positional dot product between the first attention weight map and the deep feature map of the brushless controller state multimodal encoding to obtain a first fault-specific feature map; and performing global mean pooling on the first fault-specific feature map to obtain a first fault-specific feature vector.

[0009] In one possible implementation, inputting the deep feature map of the brushless controller state multimodal encoding into a first fault attention branch to obtain a first attention weight map includes: inputting the deep feature map of the brushless controller state multimodal encoding into a convolutional layer of the first fault attention branch to obtain a single-channel attention score map, wherein the number of channels in the convolutional layer is 1; and inputting the single-channel attention score map into a sigmoid activation layer of the first fault attention branch to obtain the first attention weight map.

[0010] In one possible implementation, performing independent fault probability prediction on each fault-specific feature vector in the set of fault-specific feature vectors to obtain a fault probability vector includes: inputting each fault-specific feature vector in the set of fault-specific feature vectors into a classification head to obtain the fault probability vector composed of multiple fault probability values.

[0011] In one possible implementation, generating a multi-label diagnostic result based on the fault probability vector to obtain a composite fault diagnostic result includes: traversing each element in the fault probability vector and comparing it with a decision threshold to obtain the composite fault diagnostic result.

[0012] According to another aspect of this application, a brushless controller data analysis system is provided, comprising: a multi-source heterogeneous frequency data acquisition module for acquiring high-frequency current data and low-frequency temperature data; a multi-modal feature extraction and fusion module for extracting and merging features from the high-frequency current data and low-frequency temperature data to obtain a brushless controller state multi-modal feature map; a multi-modal deep feature encoding module for inputting the brushless controller state multi-modal feature map into a backbone network based on a convolutional neural network model to obtain a brushless controller state multi-modal encoded deep feature map; a parallel fault attention feature extraction module for inputting the brushless controller state multi-modal encoded deep feature map into a parallel fault attention branch to obtain a set of fault-specific feature vectors; an independent fault probability prediction module for performing independent fault probability prediction on each fault-specific feature vector in the set of fault-specific feature vectors to obtain a fault probability vector; and a multi-label fault diagnosis decision module for generating multi-label diagnosis results based on the fault probability vectors to obtain a composite fault diagnosis result.

[0013] Compared with existing technologies, the brushless controller data analysis method and system provided in this application firstly constructs a multimodal feature map of the brushless controller state by fusing high-frequency current data and low-frequency temperature data. Then, using a convolutional neural network, it automatically extracts the deep feature map of the brushless controller state multimodal encoding from this map. Next, it inputs the deep feature map of the brushless controller state multimodal encoding into a parallel fault attention branch to obtain a set of fault-specific feature vectors. This successfully decomposes the complex composite fault problem into multiple parallel and independent single-fault identification tasks. Finally, based on their respective specific features, independent probability predictions are performed, thereby accurately and reliably achieving multi-label composite fault diagnosis. This achieves accurate decoupling and diagnosis of composite faults in brushless controllers, significantly improving the operational reliability and safety of power systems in critical equipment. Attached Figure Description

[0014] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0015] Figure 1 A schematic flowchart illustrating a brushless controller data analysis method according to an embodiment of this application is shown.

[0016] Figure 2 The illustration shows a schematic flowchart of step S2 in the brushless controller data analysis method according to an embodiment of this application.

[0017] Figure 3 The figure shows a schematic flowchart of step S4 in the brushless controller data analysis method according to an embodiment of the present application.

[0018] Figure 4 The figure shows a schematic diagram of the model architecture of step S4 in the brushless controller data analysis method according to an embodiment of this application.

[0019] Figure 5 The figure shows a schematic flowchart of step S41 in the brushless controller data analysis method according to an embodiment of the present application.

[0020] Figure 6 The figure shows a schematic block diagram of a brushless controller data analysis system according to an embodiment of the present application. Detailed Implementation

[0021] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0022] Figure 1 A schematic flowchart illustrating a brushless controller data analysis method and system according to an embodiment of this application is shown. Figure 1 As shown, this application provides a brushless controller data analysis method, including: S1, acquiring high-frequency current data and low-frequency temperature data; S2, performing feature extraction and feature merging on the high-frequency current data and low-frequency temperature data to obtain a brushless controller state multimodal feature map; S3, inputting the brushless controller state multimodal feature map into a backbone network based on a convolutional neural network model to obtain a brushless controller state multimodal encoding deep feature map; S4, inputting the brushless controller state multimodal encoding deep feature map into a parallel fault attention branch to obtain a set of fault-specific feature vectors; S5, performing independent fault probability prediction on each fault-specific feature vector in the set of fault-specific feature vectors to obtain a fault probability vector; S6, generating multi-label diagnostic results based on the fault probability vectors to obtain a composite fault diagnostic result.

[0023] For example, in step S1, high-frequency current data and low-frequency temperature data are acquired. It should be understood that high-frequency current data can instantaneously and precisely reflect the electromagnetic stress, load fluctuations, and characteristic harmonics and current distortions caused by mechanical or electrical anomalies (such as bearing wear, rotor demagnetization, phase-to-phase short circuits, etc.) within the controller, serving as the core information carrier for capturing dynamic fault characteristics. Simultaneously, low-frequency temperature data reveals, from another dimension, the long-term thermal stress of key power components (such as MOSFETs), the effectiveness of the heat dissipation system, and the presence of slowly changing abnormal states such as continuous overload. Combining these two types of modal data with different properties and time scales allows for the construction of a far more comprehensive, three-dimensional, and robust controller state profile than a single data source, laying a solid data foundation for the subsequent accurate differentiation and diagnosis of complex composite faults.

[0024] In one embodiment, two types of data streams are synchronously acquired from the brushless controller's built-in or external sensor system via a data acquisition interface. Specifically, for high-frequency current data, current magnitude is monitored in real time by current sensors (e.g., Hall effect sensors or sampling resistors) deployed on the controller's three-phase output lines, and then converted into digital signals by a high-speed analog-to-digital converter (ADC) at a set high sampling frequency. For low-frequency temperature data, temperature is monitored by temperature sensors (e.g., thermistors or thermocouples) mounted on key heat-generating locations such as the controller's power module heatsink, and sampled at a lower frequency. Although the sampling frequencies of the two are significantly different, it is necessary to ensure that their timestamps are aligned or correlated during the acquisition process to guarantee that any current analysis window in subsequent analysis can match the effective temperature state within its corresponding time period, achieving data synchronization in the time dimension.

[0025] In one specific embodiment, acquiring high-frequency current data and low-frequency temperature data includes: acquiring the three-phase output current of the controller at a sampling frequency set to 20 kHz; and acquiring the temperature on the heatsink of the controller's power module (e.g., MOSFET) at a sampling frequency set to 1 Hz. In this example, the 20 kHz current sampling frequency is sufficient to capture most of the characteristic frequency information generated by electrical and mechanical faults, while the 1 Hz temperature sampling frequency is perfectly adequate for economically and effectively monitoring changes in the controller's thermal state.

[0026] For example, in step S2, feature extraction and feature merging are performed on high-frequency current data and low-frequency temperature data to obtain a multimodal feature map of the brushless controller state. It should be understood that the fault characteristics inherent in the original one-dimensional time-series data stream, especially the high-frequency current signal, are often hidden within complex waveform details and dynamic changes, making them difficult to utilize directly. While temperature data changes gradually, its correlation with current dynamics is crucial for diagnosis. Therefore, these raw data must be transformed into a more advanced format suitable for deep learning models (especially convolutional neural networks). By transforming the current data into a two-dimensional time-spectrum graph, its dynamic characteristics can be presented visually. Simultaneously, mapping the temperature data into a two-dimensional map and fusing it with the current map creates a unified multimodal feature map. This feature map not only retains the individual characteristics but also establishes a spatiotemporal correlation between the two, providing unprecedentedly rich information for the model to understand the controller state from a global perspective and decouple complex fault characteristics.

[0027] In one embodiment, such as Figure 2 As shown, feature extraction and feature merging are performed on high-frequency current data and low-frequency temperature data to obtain a multimodal feature map of the brushless controller state, including: S21, extracting a current analysis window from the high-frequency current data; S22, performing a short-time Fourier transform on the current analysis window to obtain a current-time spectrum; S23, calculating the mean of the low-frequency temperature data to obtain a aggregated temperature value; S24, creating a two-dimensional matrix with the same dimension as the current-time spectrum, and filling each position of the two-dimensional matrix with the aggregated temperature value to obtain a temperature distribution map; S25, merging the current-time spectrum and the temperature distribution map along the channel dimension to obtain the multimodal feature map of the brushless controller state.

[0028] Specifically, firstly, a fixed-length segment is extracted from the acquired continuous high-frequency current data as a current analysis window, for example, a time series containing 2048 sampling points. Next, a Short-Time Fourier Transform (STFT) is applied to this current analysis window. The STFT converts the one-dimensional time-domain current signal into a two-dimensional current-time spectrum. The horizontal axis of this spectrum typically represents time, and the vertical axis represents frequency. The brightness or color of each point in the spectrum represents the energy intensity of the signal at that time and frequency point, thus visualizing the dynamic characteristics of the current. Simultaneously, for the low-frequency temperature data corresponding to the current analysis window in time, the average of all temperature readings within this time period is calculated to obtain an aggregated temperature value that represents the macroscopic thermal state of the controller during that period. Then, to align the temperature characteristics with the current-time spectrum in dimension, the method creates a new matrix with the exact same two-dimensional dimensions as the current-time spectrum and fills each position of this new matrix with the single aggregated temperature value calculated in the previous step, thereby forming a uniform temperature distribution map. Finally, the current time spectrum and temperature distribution map of a single channel are merged along the channel dimension to generate a brushless controller state multimodal feature map with two channels and the same size as the original time spectrum.

[0029] For example, in step S3, the brushless controller state multimodal feature map is input into the backbone network based on a convolutional neural network model to obtain a deep feature map of the brushless controller state multimodal encoding. It should be understood that although the brushless controller state multimodal feature map is rich in information, it is essentially a relatively rudimentary, image-like representation, in which deep, abstract, and nonlinear correlation features related to faults are not explicitly revealed. To automatically and efficiently learn truly discriminative fault modes from this high-dimensional data, the powerful feature extraction capabilities of deep learning models must be utilized. Convolutional neural networks (CNNs), due to their excellent performance in image processing, are particularly suitable for processing such two-dimensional map data. Through a mature CNN backbone network, hierarchical features ranging from simple (e.g., edges, textures) to complex (e.g., specific harmonic combination modes) can be extracted layer by layer automatically from the map, ultimately generating a highly condensed deep feature representation containing rich semantic information. This replaces the cumbersome manual feature engineering in traditional methods that relies on expert knowledge, providing an ideal input for subsequent accurate fault decoupling and diagnosis.

[0030] Specifically, the multimodal feature map of the brushless controller state is fed into a convolutional neural network serving as the backbone. Inside the network, the map sequentially passes through a carefully stacked series of convolutional layers, activation function layers, and pooling layers. The convolutional layers perform sliding window operations on the input map using multiple learnable filters (or kernels) to detect specific patterns within local regions, such as specific frequency distributions or textures. Each filter generates a new feature map highlighting the location of its target pattern in the original map. Activation function layers (such as ReLU) introduce non-linearity into the network, enabling it to learn more complex data mappings. Pooling layers (such as max pooling or average pooling) downsample the feature map, reducing data size and computational complexity while also making the learned features more invariant to small changes in location. Through these layers of convolution and pooling operations, the network can automatically construct a feature hierarchy from low to high levels. Ultimately, the output layer of the backbone network will generate a deep feature map containing the richest and most abstract semantic information of the brushless controller state multimodal encoding.

[0031] In one specific embodiment, the backbone network based on the convolutional neural network model is the MobileNet model. It should be understood that MobileNetV2, as an advanced and efficient convolutional neural network model, was selected as the backbone network for feature extraction in this application because it maintains high-precision feature extraction capabilities while possessing significant advantages such as low computational cost, few parameters, and fast running speed, making it particularly suitable for deployment in industrial diagnostic scenarios with high real-time requirements. Its ingenious architecture design enables it to learn and extract deep abstract features for fault diagnosis layer by layer automatically from the input multimodal feature map of the brushless controller state.

[0032] Specifically, the core of the MobileNetV2 model architecture lies in the inverted residual module with a linear bottleneck. Unlike traditional residual modules that first compress channels, then convolution, and finally expand channels, this module adopts an inverted design of expansion-convolution-compression. Specifically, an inverted residual module consists of three key parts: first, a 1x1 expanded convolutional layer that takes a bottleneck feature map with fewer channels as input and expands its channel dimension by several times (this multiple is called the expansion factor), thus performing feature transformation in a higher-dimensional space; second, a 3x3 depthwise separable convolution, which is responsible for efficient spatial filtering on the expanded high-dimensional feature map; and finally, a 1x1 projective convolutional layer that recompresses the high-dimensional feature map back to a bottleneck state with fewer channels. Crucially, this final projective layer uses linear activation (i.e., it does not use non-linear activation functions such as ReLU) to avoid information loss in the low-dimensional representation. Furthermore, when the input and output dimensions are the same, the two ends of the module are directly connected through a residual connection, which facilitates gradient propagation and allows the network to be built deeper. The entire MobileNetV2 backbone network is formed by an ordered stacking of a series of inverted residual modules after an initial standard convolutional layer. The depth-separable convolutional stride of some modules is set to 2 to achieve a gradual reduction in the size of the feature map space (i.e., downsampling).

[0033] In a specific embodiment, taking a brushless controller state multimodal feature map as an example, which is a data tensor of size (128, 128, 2), where 128x128 is the spatial resolution of the map and 2 represents that it contains two channels: current and temperature. The processing flow of this data tensor in the MobileNetV2 backbone network can be summarized as follows: First, the (128, 128, 2) input map passes through a standard 3x3 convolutional layer at the beginning of the network. This layer typically has a stride of 2 and expands the number of channels to an initial value, such as 32. After this layer, the data tensor becomes (64, 64, 32), with the spatial size halved and the channel dimension increased. Next, this (64, 64, 32) feature map enters a series of stacked inverted residual modules for processing. For example, it flows through the first inverted residual module, which has a scaling factor of 6 and outputs 24 channels. Inside the module, the data is first expanded to (64, 64, 192) by a 1x1 convolution, then passed through a 3x3 depthwise separable convolution (stride 1), maintaining the size (64, 64, 192), and finally projected back to (64, 64, 24) by a 1x1 linear convolution. Subsequently, the feature map continues to flow through subsequent inverted residual modules. When encountering a module with a stride of 2, the spatial size of the data is halved again. For example, when the (64, 64, 24) feature map enters a module with a stride of 2 and 32 output channels, its internal 3x3 depthwise separable convolution will be performed with a stride of 2, thus downsampling the expanded (64, 64, 144) feature map to (32, 32, 144), and the final output feature map size becomes (32, 32, 32). This process repeats continuously, with the data tensor propagating forward through the network, its spatial size being gradually compressed, while its channel dimensions are adjusted according to the network design, resulting in increasingly higher levels of feature abstraction and semantics. Ultimately, after the data flows through all modules of the backbone network, a final output is obtained that is highly compressed spatially but contains extremely rich and abstract feature information in its channels. In the embodiments of this application, the size of the deep feature map of the brushless controller state multimodal encoding of this final output is (8, 8, 256).

[0034] For example, in step S4, the deep feature map of the brushless controller state multimodal encoding is input into a parallel fault attention branch to obtain a set of fault-specific feature vectors. It should be understood that, because the deep feature map of the brushless controller state multimodal encoding contains both shared and distinguishable features relative to various faults, standard attention mechanisms may suffer from attention dispersion or ambiguous feature attribution. For instance, when controller overload and bearing wear faults occur concurrently, a strong shared feature (such as overall current distortion caused by drastic load changes) may receive high attention in all attention branches, completely ignoring or masking specific features that truly indicate a weak fault (such as weak sideband harmonics generated by early bearing wear). Therefore, this application proposes that the attention of a branch should not only focus on features that benefit its own task but also actively suppress shared features that also benefit other potential tasks (i.e., other faults), thereby forcing the attention weight map to learn truly distinguishable specific features. Based on this, by inputting the deep feature map of the brushless controller state multimodal encoding into the parallel fault attention branch, the unique and pure feature representation of each potential fault can be accurately decoupled from the mixed deep features.

[0035] Specifically, the parallel fault attention branch is a specially designed network structure that creates an independent, parallel computational path or branch for each type of fault requiring diagnosis. All these branches receive the exact same input—deep feature maps from the backbone network—and process them in parallel. The final output of this step is a set of fault-specific feature vectors, which is a collection of multiple one-dimensional vectors. Each vector is a compact feature representation of a specific fault, extracted from its corresponding fault attention branch, and theoretically, has minimized interference from other fault information.

[0036] In one embodiment, such as Figure 3 and Figure 4 As shown, the brushless controller state multimodal encoding deep feature map is input into a parallel fault attention branch to obtain a set of fault-specific feature vectors, including: S41, inputting the brushless controller state multimodal encoding deep feature map (e.g., as...) into a parallel fault attention branch to obtain a set of fault-specific feature vectors. Figure 4 The input of Fa (as shown) to the first fault attention branch (e.g., as shown) Figure 4 The CNN shown is used to obtain the first attention weight map (e.g., as shown in the diagram). Figure 4 S42, calculate the position-based dot multiplication between the first attention weight map and the brushless controller state multimodal coding deep feature map to obtain the first fault-specific feature map (e.g., as shown in Fb); ... Figure 4S43, perform global mean pooling on the first fault-specific feature map to obtain the first fault-specific feature vector (e.g., as shown in Fc); ... Figure 4 The Vc shown is shown.

[0037] In other words, the same logic is applied to each parallel fault branch. For any fault branch (taking the first fault attention branch as an example; it should be noted that "first" here is only for illustrative purposes and can refer to any fault branch), the goal is to generate a first fault-specific feature vector. This process can be broken down into three core sub-steps: generating an attention weight map, generating a fault-specific feature map, and generating a fault-specific feature vector.

[0038] In one embodiment, such as Figure 5 As shown, the process of inputting the deep feature map of the brushless controller state multimodal encoding into the first fault attention branch to obtain the first attention weight map includes: S411, inputting the deep feature map of the brushless controller state multimodal encoding into the convolutional layer of the first fault attention branch to obtain a single-channel attention score map, wherein the number of channels in the convolutional layer is 1; S412, inputting the single-channel attention score map into the sigmoid activation layer of the first fault attention branch to obtain the first attention weight map. Then, this complete process is repeated for all parallel fault attention branches to obtain a set containing all fault-specific feature vectors.

[0039] To illustrate more specifically, in one particular embodiment, three types of faults to be diagnosed (F1 - bearing wear, F2 - rotor demagnetization, F3 - controller overload) were processed. Specifically, in this embodiment, the parallel fault attention branch was configured with three parallel fault attention branches. The deep feature map of the brushless controller state multimodal encoding was simultaneously input into these three branches, and the following operations were performed on each branch (taking the first fault attention branch F1 as an example): First, the input deep feature map was processed using independent convolution operations to obtain a single-channel attention score. Next, the single-channel attention score was input into the Sigmoid activation layer of the first fault attention branch F1, and its value was normalized to the (0, 1) interval, finally obtaining the first attention weight map. This weight map highlighted the area in the deep feature map most relevant to the F1 fault (bearing wear). Then, the first attention weight map was multiplied by position with the deep feature map of the brushless controller state multimodal encoding. This is equivalent to using the weight map as a sieve to filter out irrelevant information and obtain the first fault-specific feature map. Perform global mean pooling on the first fault-specific feature map, compressing it from a two-dimensional map into a one-dimensional vector, i.e., the first fault-specific feature vector. Perform the same operation on branches F2 and F3, ultimately obtaining a set containing three fault-specific feature vectors.

[0040] Since the deep feature map of the multimodal coding of the brushless controller state contains features shared with various faults and features that are distinct, the standard attention mechanism may suffer from problems of attention dispersion or feature attribution ambiguity. Therefore, it is believed that the attention of a certain branch should not only focus on features that are beneficial to its own task, but also actively suppress those shared features that are also beneficial to other potential tasks (i.e. other faults), thereby forcing the attention weight map to learn truly distinctive specific features.

[0041] In a preferred embodiment, inputting the deep feature map of the brushless controller state multimodal encoding into a first fault attention branch to obtain a first attention weight map includes: performing a convolution operation on the deep feature map of the brushless controller state multimodal encoding to obtain prototype feature maps corresponding to the first fault attention branch and other fault attention branches respectively; calculating a cross-suppression map based on the prototype feature maps of other fault attention branches besides the first fault attention branch; performing cross-attention suppression calculation based on the prototype feature map of the first fault attention branch and the cross-suppression map to generate a corrected prototype feature map; inputting the corrected prototype feature map into the convolutional layer of the first fault attention branch to obtain a single-channel attention score map, and passing the single-channel attention score map through an activation layer to obtain the first attention weight map.

[0042] Specifically, for the deep feature map of the state multimodal encoding of the brushless controller, for example, denoted as... Firstly, regarding the first Apply convolution operation to faulty branches Obtain the corresponding prototype feature map, i.e. And calculate the cross-suppression graph representation, i.e.: ;in, It is the number of faulty branches. Indicates the first Prototype feature map of the faulty branch. Indicates all Summing the faulty branches, This indicates addition by position. Indicates the first The cross-suppression plot of the faulty branch represents the degree of competition for attention among other branches, that is, the average degree of feature attention of other branches besides this branch.

[0043] Then, based on the cross-suppression diagram representation, it is possible to pass through the temperature coefficient... The function implements a soft cross-attention inhibition process, namely: ;here, Indicates the first Prototype feature map of the faulty branch. This indicates subtraction by position. Indicates the first Cross-suppression plot of faulty branches, This represents element-wise multiplication, also known as positional dot product, which involves adding the corresponding value to each element at each position in the feature map. express The reciprocal of the variance, Represents an sigmoid function. This represents the corrected prototype feature map. Indicates the first The fault has a specific focus on the feature map distribution, such as a location within the feature map, where a positive difference indicates the fault's position. A negative value indicates that this area is more relevant than the average value of other faults; conversely, a negative value suggests that this area is more like a shared characteristic or a characteristic of other faults. Meanwhile, Cross-suppression diagram representation The reciprocal of the variance, which serves as The temperature coefficient of the function. That is, if... Very small, then The function becomes very steep, approaching a step function, thus enhancing the effectiveness of specific attention when shared attention is high. Conversely, if A large number indicates that the degree of shared attention among the various faulty branches is not high, which can make the effect of dedicated attention smoother and gentler.

[0044] Then, the corrected prototype feature map can be... Enter the first The convolutional layers of the fault attention branch are used to obtain a single-channel attention score map, where the number of channels in the convolutional layer is 1. This allows for the introduction of competitive suppression relationships in each parallel branch, enabling each branch to reinforce specific features rather than shared features (e.g., an overall increase in signal energy may simultaneously indicate overload, short circuit, or demagnetization). This decouples anomalies corresponding to different faults in complex fault scenarios while counteracting feature masking. Specifically, when a strong shared feature (such as current distortion caused by drastic load changes) appears, it may receive high attention in all attention branches, thus ignoring specific features that truly indicate a weak fault (such as sideband harmonics in early bearing wear).

[0045] For example, in step S5, each fault-specific feature vector in the set of fault-specific feature vectors is independently predicted to obtain a fault probability vector. It should be understood that since each fault-specific feature vector is designed to contain only information strongly correlated with a specific fault, independently predicting them is logically necessary for achieving multi-label composite fault diagnosis. This independence ensures that the judgment of the probability of occurrence of one fault (such as bearing wear) is not affected by the strength or presence of another concurrent fault (such as controller overload), thereby decomposing the complex, intertwined composite fault problem into a series of parallel, simple binary judgment problems, greatly improving the clarity and accuracy of the diagnosis.

[0046] In one embodiment, independently predicting the fault probability of each fault-specific feature vector in the set of fault-specific feature vectors to obtain a fault probability vector includes: inputting each fault-specific feature vector in the set of fault-specific feature vectors into a classification head to obtain the fault probability vector composed of multiple fault probability values. Specifically, first, a fault-specific feature vector is taken from the set of fault-specific feature vectors. Then, this vector is input into its corresponding classification head, which is specifically trained for this type of fault. In a specific embodiment, the classification head is a fully connected layer that performs a linear transformation on the input feature vector, mapping it to a new space. Next, the result of the linear transformation is input into an activation function, typically the Sigmoid function. The key function of the Sigmoid function is to compress its input values ​​into the interval (0, 1), and this output value can be naturally interpreted as the probability of the fault occurring. This process generates a single probability value for the fault-specific feature vector. Finally, the above operation is repeated independently and in parallel for each fault-specific feature vector in the set, with each vector calculating a probability value through its own classification head. Combining all these probability values ​​according to a predetermined fault sequence forms the final fault probability vector.

[0047] For example, in step S6, multi-label diagnostic results are generated based on the fault probability vector to obtain composite fault diagnostic results. It should be understood that by generating multi-label diagnostic results from the fault probability vector, it is possible to ultimately determine whether a single fault exists, or whether a composite fault situation with multiple faults exists simultaneously, thereby achieving accurate identification and reporting of concurrent faults.

[0048] In one embodiment, generating a multi-label diagnostic result based on the fault probability vector to obtain a composite fault diagnostic result includes: traversing each element in the fault probability vector and comparing it with a decision threshold to obtain the composite fault diagnostic result. Specifically, this process first requires setting a reasonable decision threshold. Then, each probability value in the fault probability vector is checked sequentially. For each probability value, it is compared with the preset decision threshold. If the currently checked probability value is greater than the decision threshold, the system determines that the corresponding specific fault type exists or has occurred; conversely, if the probability value is less than or equal to the decision threshold, the fault is determined to be non-existent or not occurred. This independent comparison and judgment process is repeated for all elements in the vector until each potential fault has a clear yes or no determination. Finally, all fault types determined to exist are integrated to form a final, comprehensive diagnostic statement, i.e., the "composite fault diagnostic result".

[0049] In one specific embodiment, firstly, a decision threshold is set, for example, 0.6; this is merely an example. Next, each element in the fault probability vector [0.95, 0.12, 0.88] obtained in the previous step is iterated and compared with the decision threshold: F1 (bearing wear): 0.95 > 0.6, indicating the presence of this fault. F2 (rotor demagnetization): 0.12 < 0.6, indicating the absence of this fault. F3 (controller overload): 0.88 > 0.6, indicating the presence of this fault. Finally, all the judgment results are combined to generate the final composite fault diagnosis result: that is, the brushless controller has a composite fault of bearing wear and controller overload.

[0050] In summary, the brushless controller data analysis method provided in this application first constructs a multimodal feature map of the brushless controller state by fusing high-frequency current data and low-frequency temperature data. Then, using a convolutional neural network, it automatically extracts the deep feature map of the brushless controller state multimodal encoding from this multimodal map. Next, it inputs the deep feature map of the brushless controller state multimodal encoding into a parallel fault attention branch to obtain a set of fault-specific feature vectors. This successfully decomposes the complex composite fault problem into multiple parallel and independent single fault identification tasks. Finally, based on their respective specific features, independent probability prediction is performed, thereby accurately and reliably achieving multi-label composite fault diagnosis. In this way, accurate decoupling and diagnosis of composite faults in brushless controllers are achieved, significantly improving the operational reliability and safety of the power system in key equipment.

[0051] This application also provides a brushless controller data analysis system for performing the above-described brushless controller data analysis method, such as... Figure 6 As shown, the brushless controller data analysis system 600 includes: a multi-source heterogeneous frequency data acquisition module 610, used to acquire high-frequency current data and low-frequency temperature data; a multi-modal feature extraction and fusion module 620, used to extract and merge features from the high-frequency current data and low-frequency temperature data to obtain a brushless controller state multi-modal feature map; a multi-modal deep feature encoding module 630, used to input the brushless controller state multi-modal feature map into a backbone network based on a convolutional neural network model to obtain a brushless controller state multi-modal encoding deep feature map; a parallel fault attention feature extraction module 640, used to input the brushless controller state multi-modal encoding deep feature map into a parallel fault attention branch to obtain a set of fault-specific feature vectors; an independent fault probability prediction module 650, used to perform independent fault probability prediction on each fault-specific feature vector in the set of fault-specific feature vectors to obtain a fault probability vector; and a multi-label fault diagnosis decision module 660, used to generate multi-label diagnosis results based on the fault probability vectors to obtain a composite fault diagnosis result.

[0052] This application also provides a computer-readable storage medium storing computer program code. When the computer program code is run on a computer, the computer executes the above-described related method steps to implement the brushless controller data analysis method provided in the above embodiments.

[0053] This application also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned related steps to implement the brushless controller data analysis method provided in the above embodiments.

[0054] In this application, the system, computer-readable storage medium, or computer program product provided in the embodiments are all used to execute the corresponding methods provided above. Therefore, the beneficial effects that can be achieved can be referred to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

[0055] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments.

[0056] The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous. The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A brushless controller data analysis method, characterized in that, include: Acquire high-frequency current data and low-frequency temperature data; Feature extraction and feature merging were performed on high-frequency current data and low-frequency temperature data to obtain a multimodal feature map of the brushless controller state. The state multimodal feature map of the brushless controller is input into the backbone network based on the convolutional neural network model to obtain the deep feature map of the state multimodal encoding of the brushless controller. The deep feature map of the brushless controller state multimodal encoding is input into the parallel fault attention branch to obtain a set of fault-specific feature vectors; Independent fault probability prediction is performed on each fault-specific feature vector in the set of fault-specific feature vectors to obtain a fault probability vector; Multi-label diagnostic results are generated based on the fault probability vector to obtain composite fault diagnostic results.

2. The brushless controller data analysis method according to claim 1, characterized in that, Feature extraction and feature merging are performed on high-frequency current data and low-frequency temperature data to obtain a multimodal feature map of the brushless controller state, including: Extract a current analysis window from the high-frequency current data; A short-time Fourier transform is performed on the current analysis window to obtain the current time-frequency spectrum. The average value of the low-frequency temperature data is calculated to obtain the polymerization temperature value; Create a two-dimensional matrix with the same dimensions as the current-time spectrum, and fill the various positions of the two-dimensional matrix with the aggregated temperature values ​​to obtain a temperature distribution map; The current-time spectrum and the temperature distribution map are merged along the channel dimension to obtain the state multimodal feature map of the brushless controller.

3. The brushless controller data analysis method according to claim 1, characterized in that, The backbone network of the convolutional neural network model is the MobileNet model.

4. The brushless controller data analysis method according to claim 1, characterized in that, The deep feature map of the brushless controller's state multimodal encoding is input into a parallel fault attention branch to obtain a set of fault-specific feature vectors, including: The deep feature map of the brushless controller state multimodal encoding is input into the first fault attention branch to obtain the first attention weight map; The first fault-specific feature map is obtained by calculating the position-based multiplication between the first attention weight map and the deep feature map of the brushless controller state multimodal coding. The first fault-specific feature map is subjected to global mean pooling to obtain the first fault-specific feature vector.

5. The brushless controller data analysis method according to claim 4, characterized in that, The deep feature map of the brushless controller state multimodal encoding is input into the first fault attention branch to obtain the first attention weight map, including: The deep feature map of the brushless controller state multimodal encoding is input into the convolutional layer of the first fault attention branch to obtain a single-channel attention score map, wherein the number of channels of the convolutional layer is 1; The single-channel attention score map is input into the Sigmoid activation layer of the first fault attention branch to obtain the first attention weight map.

6. The brushless controller data analysis method according to claim 1, characterized in that, Independent fault probability prediction is performed on each fault-specific feature vector in the set of fault-specific feature vectors to obtain a fault probability vector, including: Each fault-specific feature vector in the set of fault-specific feature vectors is input into the classification head to obtain the fault probability vector composed of multiple fault probability values.

7. The brushless controller data analysis method according to claim 1, characterized in that, The process of generating multi-label diagnostic results based on the fault probability vector to obtain composite fault diagnostic results includes: traversing each element in the fault probability vector and comparing it with a decision threshold to obtain the composite fault diagnostic results.

8. A brushless controller data analysis system, characterized in that, include: A multi-source heterogeneous frequency data acquisition module is used to acquire high-frequency current data and low-frequency temperature data; The multimodal feature extraction and fusion module is used to extract and merge features from high-frequency current data and low-frequency temperature data to obtain the state multimodal feature map of the brushless controller; The multimodal deep feature encoding module is used to input the multimodal feature map of the brushless controller state into the backbone network based on the convolutional neural network model to obtain the multimodal encoded deep feature map of the brushless controller state. The parallel fault attention feature extraction module is used to input the deep feature map of the brushless controller state multimodal encoding into the parallel fault attention branch to obtain a set of fault-specific feature vectors; An independent fault probability prediction module is used to perform independent fault probability prediction on each fault-specific feature vector in the set of fault-specific feature vectors to obtain a fault probability vector. The multi-label fault diagnosis decision module is used to generate multi-label diagnosis results based on the fault probability vector to obtain composite fault diagnosis results.

Images