Real-time diagnosis method and system for amorphous strip invisible defects

Abstract

This application provides a real-time diagnostic method and system for hidden defects in amorphous ribbon, belonging to the field of non-destructive testing technology for materials. The method includes: simultaneously acquiring at least two physical signals from the amorphous ribbon using a multi-physics synchronous sensing array; extracting multiple feature parameters from the signals; inputting the feature parameters into a pre-trained machine learning diagnostic model, and having the model output a diagnostic result containing defect information. The system includes: the multi-physics synchronous sensing array; and a processing unit connected to the sensing array and configured to execute the above method. By fusing multiple physical signals and utilizing a machine learning model for analysis, this application solves the problems of blind spots and inability to perform real-time, high-precision diagnosis in existing single-detection methods, enabling comprehensive, online, and high-precision diagnosis of hidden defects in amorphous ribbon, thereby improving product quality control.

Description

Technical Field

[0001] This application relates to the field of nondestructive testing technology for materials, and in particular to a method and system for real-time diagnosis of hidden defects in amorphous ribbons. Background Technology

[0002] Amorphous ribbon is a high-performance soft magnetic material, and its quality is crucial to the final performance of electronic components and power equipment. During the production of amorphous ribbon, fluctuations in process parameters such as cooling rate, compositional uniformity, and ribbon tension can lead to various "hidden" defects, such as microcracks, porosity, compositional segregation, and uneven internal stress. These defects are characterized by their tiny size or location within the material, making them difficult to identify effectively and comprehensively using traditional single-method detection.

[0003] Currently, quality inspection methods for amorphous ribbon mainly rely on a single physical principle. For example, eddy current testing is used solely to detect changes in surface conductivity, thus inferring the presence of cracks; or optical inspection is used solely to identify surface scratches, dirt, and other defects. The limitation of these methods lies in the existence of "detection blind spots," meaning that defect types insensitive to certain physical principles may be missed. For instance, eddy current testing struggles to detect stress concentration areas within the material, while optical inspection cannot detect internal compositional segregation. Furthermore, many existing inspection methods require offline operation, failing to meet the demands of modern industrial production for real-time online monitoring. This often results in a large number of defective products already produced by the time quality problems are discovered, and no effective data is available for real-time adjustments to production processes, leading to significant economic losses and resource waste. While general approaches combining multiple sensors with machine learning models have emerged in other technological fields, there remains a lack of effective solutions for creatively applying this approach to the specific industrial inspection scenario of amorphous ribbon and specifically addressing the technical challenges of online, real-time, and high-precision diagnosis of its hidden defects. Summary of the Invention

[0004] The purpose of this application is to provide a real-time diagnostic method and system for hidden defects in amorphous ribbons, so as to solve the problem that the existing technology uses a single detection method, which has a detection blind spot and cannot perform comprehensive, real-time and high-precision online diagnosis of hidden defects in the production process of amorphous ribbons.

[0005] To address the aforementioned technical problems, this application provides a real-time diagnostic method for hidden defects in amorphous ribbon, comprising the following steps: synchronously acquiring at least two physical signals of the amorphous ribbon using a multi-physics synchronous sensing array; extracting multiple pre-selected feature parameters from the at least two physical signals; inputting the feature parameters into a pre-trained machine learning diagnostic model, and having the machine learning diagnostic model output a diagnostic result containing defect information of the amorphous ribbon, wherein the defect information includes at least one of the following: whether the defect exists, defect location information, and defect type information.

[0006] Optionally, the at least two physical signals include a magnetic field signal and a thermal field signal.

[0007] Optionally, the at least two physical signals include eddy current signals and acoustic signals.

[0008] Furthermore, the machine learning diagnostic model is a convolutional neural network model or a recurrent neural network model.

[0009] This application also provides a real-time diagnostic system for hidden defects in amorphous ribbon, comprising: a multi-physics synchronous sensing array configured to synchronously acquire at least two physical signals of the amorphous ribbon; and a processing unit connected to the multi-physics synchronous sensing array, wherein the processing unit stores a pre-trained machine learning diagnostic model; the processing unit is configured to: extract multiple pre-selected feature parameters from the at least two physical signals; and input the feature parameters into the machine learning diagnostic model to generate a diagnostic result containing defect information of the amorphous ribbon, wherein the defect information includes at least one of the following: whether the defect exists, defect location information, and defect type information.

[0010] In a preferred embodiment of this application, the processing unit is a single integrated diagnostic host.

[0011] In another preferred embodiment of this application, the processing unit includes an edge computing unit and a diagnostic server.

[0012] Furthermore, the edge computing unit is configured to extract the feature parameters from the at least two physical signals; and the diagnostic server is configured to receive the feature parameters extracted by the edge computing unit and run the machine learning diagnostic model to generate the diagnostic results.

[0013] Optionally, the multi-physics synchronous sensing array includes a Hall sensor array and an infrared thermal imager.

[0014] Optionally, the multiphysics synchronous sensing array includes an eddy current probe array and an acoustic sensor.

[0015] Compared with existing technologies, the real-time diagnosis method and system for hidden defects in amorphous ribbons provided in this application have the following advantages: First, by fusing information from at least two physical signals through a multi-physics synchronous sensing array, and by leveraging the complementary sensitivity of different signals to different defects, the blind spots of single detection methods are effectively overcome, significantly improving the detection rate of various hidden defects and achieving a more comprehensive diagnosis.

[0016] Secondly, the entire solution of this invention can be integrated into the production line to realize online real-time detection of amorphous ribbon. The diagnostic results can be fed back immediately so as to adjust the process parameters immediately, avoid the generation of large-scale waste, support the closed-loop optimization of the production process, and provide more timely feedback.

[0017] Furthermore, by introducing pre-trained machine learning diagnostic models, it is possible to deeply mine defect features from complex and weak multi-source signals, achieve high-precision identification and classification of hidden defects, and even quantitatively assess the severity of defects, making the diagnostic results more reliable. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the structure of the real-time diagnostic system for hidden defects in amorphous ribbon provided in the embodiments of this application; Figure 2 A flowchart illustrating the real-time diagnosis method for hidden defects in amorphous ribbons provided in this application embodiment; Figure 3 A schematic diagram of a diagnostic system using a distributed processing architecture is provided for another embodiment of this application; Figure 4 This is a timing diagram of the signaling interaction between components in the distributed processing architecture provided in the embodiments of this application.

[0020] Key reference numerals: 10-Amorphous ribbon; 20-Multiphysics synchronous sensing array; 21-Sensor A; 22-Sensor B; 30-Processing unit; 31-Edge computing unit; 32-Diagnostic server; 40-Network. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the specific embodiments described herein are merely illustrative of the application and are not intended to limit the scope of protection of this application.

[0022] Example 1 This application provides a real-time diagnostic system and method for hidden defects in amorphous ribbon based on a centralized processing architecture. Specifically, this solution aims to achieve comprehensive, real-time online detection of magnetic property inhomogeneities and internal stress defects that may occur during the production of amorphous ribbon by integrating two physical field signals: magnetic field and thermal field.

[0023] Reference Figure 1 This is a schematic diagram of the overall structure of a real-time diagnostic system for hidden defects in amorphous ribbon provided in an embodiment of this application. The system is deployed on an amorphous ribbon production line, for example, after the rapid cooling and initial tension adjustment processes. The system mainly includes a multiphysics synchronous sensing array 20 and a processing unit 30. During production, continuous amorphous ribbon 10 passes beneath the multiphysics synchronous sensing array 20 at a preset constant speed (e.g., 10-30 m / s).

[0024] In one embodiment of this application, the multiphysics synchronous sensing array 20 is specifically composed of two different types of sensors: a Hall sensor array as sensor A 21 and an infrared thermal imager as sensor B 22. The Hall sensor array is a magnetic sensor array that uniformly arranges multiple (e.g., 64 or 128) highly sensitive Hall elements along the width direction of the amorphous ribbon 10 to acquire the leakage magnetic field signal on the surface of the amorphous ribbon 10 in real time. It is understood that, as a soft magnetic material, the stress distribution, crystallinity, or compositional segregation within the amorphous ribbon can cause changes in local permeability, resulting in an abnormal leakage magnetic field externally. Correspondingly, the infrared thermal imager, as a non-contact temperature measurement device, can simultaneously acquire a two-dimensional surface temperature distribution image of the amorphous ribbon 10 at the same detection location. Compositional segregation or microstructural inhomogeneity within the material may lead to differences in its heat dissipation characteristics during cooling, resulting in abnormal temperature distributions such as hot spots or cold spots on the surface.

[0025] To achieve synchronous signal acquisition, the Hall sensor array and the infrared thermal imager are physically precisely aligned to ensure that they detect identical transverse cross-sectional areas on the amorphous ribbon 10 at the same time. Simultaneously, their signal acquisition actions are triggered by a unified synchronization clock signal, guaranteeing strict consistency in timestamps between the magnetic field signal data stream and the thermal field image data stream. This dual synchronization in both space and time constitutes a crucial prerequisite for subsequent effective data fusion and analysis.

[0026] In this embodiment, the processing unit 30 can be specifically configured as a high-performance integrated diagnostic host, such as an industrial computer equipped with a high-performance central processing unit, a graphics processing unit, and a large-capacity memory. This integrated diagnostic host is electrically connected to the multi-physics synchronous sensor array 20 via a high-speed, low-latency data interface (e.g., Fibre Channel or 10 Gigabit Industrial Ethernet), and is responsible for receiving and processing the massive amounts of raw data transmitted from the sensor terminals. The integrated diagnostic host deploys a complete set of software algorithms, including but not limited to a data preprocessing module, a feature extraction module, and a pre-trained machine learning diagnostic model.

[0027] The following is combined Figure 2 The flowchart shown illustrates the working process of this embodiment.

[0028] In step S101, at least two physical signals of the amorphous ribbon are simultaneously acquired using a multi-physics synchronous sensing array. Specifically, as the amorphous ribbon 10 continuously passes through the detection area, the Hall sensor array (sensor A 21) continuously acquires leakage magnetic field strength values ​​distributed along the width direction of the ribbon at an extremely high sampling frequency (e.g., 100 kHz), forming a series of magnetic field signal time sequences. Simultaneously, an infrared thermal imager (sensor B 22) synchronously captures thermal images of the ribbon surface at a certain frame rate (e.g., 100 Hz), forming a thermal image sequence. These two heterogeneous signal streams are transmitted to the processing unit 30 in real time.

[0029] Subsequently, in step S102, preset feature parameters are extracted from the acquired signal. Upon receiving the synchronization signal stream, the processing unit 30 can first perform necessary preprocessing, such as high-pass filtering of the magnetic field signal to eliminate low-frequency interference from the geomagnetic field, and noise filtering and brightness correction of the thermal image. Correspondingly, the feature extraction module begins operation. For a one-dimensional magnetic field signal, this module calculates a series of time-domain and frequency-domain feature parameters. Time-domain feature parameters may include: the signal's mean, variance, peak-to-peak value, root mean square value, kurtosis (characterizing the sharpness of the signal waveform), and skewness (characterizing the symmetry of the signal waveform). Frequency-domain feature parameters are obtained by performing a fast Fourier transform on the signal and may include: the energy proportion of a specific frequency band, the dominant frequency, and the spectral entropy. These parameters can quantify the fluctuation of the leakage magnetic field signal from different dimensions. For two-dimensional thermal images, this module extracts a series of spatial and statistical feature parameters, such as: the average temperature, highest temperature, lowest temperature, temperature standard deviation, temperature gradient reflecting the degree of temperature change (which can be obtained by calculating the modulus of the temperature difference between adjacent pixels), and geometric features such as area, perimeter, and circularity of high-temperature or low-temperature regions (i.e., hot spots or cold spots) identified by the image segmentation algorithm.

[0030] In step S103, the extracted feature parameters are input into a pre-trained machine learning diagnostic model. In this embodiment, considering that the input features contain spatial information from the thermal image, a convolutional neural network model is preferred. Before inputting into the model, all feature parameters extracted from the magnetic field signal and thermal field image (e.g., 15 magnetic field features and 20 thermal field features) are combined into a fixed-length multidimensional feature vector, which comprehensively describes the multiphysics state of the amorphous ribbon 10 at the current detection location. It is understood that this convolutional neural network model is pre-trained using a large amount of labeled sample data. The training dataset contains tens of thousands of samples, each corresponding to a feature vector and its true defect label (e.g., normal, first-order stress defect, second-order compositional segregation defect, etc.), which are typically obtained through offline, more precise destructive or non-destructive analysis of the samples (e.g., metallographic analysis, X-ray diffraction). Through optimization using algorithms such as backpropagation and gradient descent, the model learns a complex nonlinear mapping relationship from a specific feature vector combination to a specific defect type.

[0031] Finally, in step S104, the defect diagnosis result is output. During the real-time diagnosis phase, the processing unit 30 inputs the real-time generated feature vector into a pre-trained convolutional neural network model. The model performs forward propagation calculations on the input feature vector, and its output layer provides a classification result. This diagnosis result can contain rich defect information; for example, it first determines whether a defect exists; if it exists, it further provides defect type information, such as "Level 1 stress defect" or "Level 2 component segregation defect." Simultaneously, it can also output defect location information, which can be precisely correlated with encoder readings and sensor positions to pinpoint the longitudinal and transverse coordinates of the defect on the entire roll of strip. Furthermore, the model's output can also be confidence scores for various types of defects, such as {"Normal": 0.95, "Stress Defect": 0.04, "Component Segregation": 0.01}, providing operators with a basis for judging the reliability of the diagnosis result. The diagnosis result is displayed in real-time on the monitoring interface of the integrated diagnostic host and can be sent to the main control system of the production line via industrial protocols (such as OPC UA) for alarm purposes or to trigger subsequent marking, rejection, and other actions.

[0032] The technical solution of this embodiment can identify magnetic property fluctuations caused by uneven internal stress and local abnormal temperature rises caused by component segregation in real time and accurately. It realizes a comprehensive diagnosis of two key hidden defects that cannot be effectively detected by a single sensor at the same time, thereby greatly improving the product quality control level.

[0033] Example 2 Another embodiment of this application provides a real-time diagnostic system for hidden defects in amorphous ribbon using a distributed processing architecture. As an optional implementation, this scheme can be considered a variation of the processing unit in Embodiment 1, and is particularly suitable for scenarios with extremely high production line speeds or ultra-wide amorphous ribbon widths, because such scenarios generate extremely large amounts of raw sensor data, posing a significant challenge to the computing power and data transmission bandwidth of the central processor. The distributed architecture effectively alleviates these problems by deploying edge computing capabilities on the production floor.

[0034] Figure 3 This is a schematic diagram of the diagnostic system architecture using a distributed processing architecture in this embodiment. Figure 1 Unlike the centralized structure, the processing unit function in this embodiment is decomposed into two physically separate entities: an edge computing unit 31 and a diagnostic server 32.

[0035] Edge computing unit 31 is typically a compact, robust, fanless embedded industrial computer installed close to the multiphysics synchronous sensor array 20 in the production environment. Its hardware configuration emphasizes high data throughput and real-time signal processing capabilities. Edge computing unit 31 connects directly to the multiphysics synchronous sensor array 20 via a short-distance, high-bandwidth dedicated cable to receive the raw physical signals.

[0036] The diagnostic server 32 is a high-performance server located in the central control room or data center, possessing powerful general-purpose computing capabilities and massive storage space. The diagnostic server 32 is connected to the edge computing unit 31 in the field via a stable and reliable industrial Ethernet (i.e., network 40).

[0037] It is understood that the configuration of the multi-physics synchronous sensing array 20 can be the same as that of Embodiment 1. For example, it can also be composed of a Hall sensor array and an infrared thermal imager, used to synchronously acquire the magnetic field signal and thermal field signal of the amorphous ribbon 10.

[0038] The workflow of this embodiment can still be followed logically. Figure 2 The steps are shown, but the entities performing each step differ. The following will refer to... Figure 4 The timing diagram shown illustrates the specific working process and signaling interaction.

[0039] In the real-time detection loop, the multi-physics synchronous sensing array 20 continuously acquires raw signals and sends high-frequency magnetic field signal data streams and thermal image video streams to the edge computing unit 31 in real time through its interface. This process corresponds to... Figure 4 The message "Acquire and send raw signal" in the middle.

[0040] After receiving the raw signal, the edge computing unit 31 has its built-in software responsible for executing... Figure 2 Steps S101 (partial) and S102 in the process. Specifically, the edge computing unit 31 first buffers and aligns the received raw signal to ensure the integrity and synchronization of the data frame. Then, it executes the same preprocessing and feature extraction algorithms described in Embodiment 1 to convert the high-dimensional, high-bandwidth raw signal stream (e.g., hundreds of megabytes per second) into a low-dimensional, low-bandwidth feature data stream (e.g., only a few gigabytes per second) in real time. This process corresponds to Figure 4 The "extract feature parameters" action in the process. After feature extraction is completed, the edge computing unit 31 packages the feature vector containing a series of feature parameters into a diagnostic request message and sends it to the remote diagnostic server 32 via the network 40. This process corresponds to... Figure 4 The "Send Feature Data (Diagnostic Request)" message.

[0041] As the core of the backend processing, diagnostic server 32's main responsibility is to execute... Figure 2 Steps S103 and S104 in the process. One or more pre-trained machine learning diagnostic models are deployed on the diagnostic server 32. Since the server has ample computing resources, a more complex model than that in Example 1 can be used here, such as an ensemble learning model that fuses a convolutional neural network for processing thermal image features, a long short-term memory network for processing magnetic field temporal features, and a gradient boosting decision tree model for final decision-making. Such complex models typically achieve higher diagnostic accuracy. When the diagnostic server 32 receives feature data from the edge computing unit 31, it inputs the feature vector into the diagnostic model for calculation. This process corresponds to... Figure 4 The "Run the model to perform diagnosis" action in the process.

[0042] After the model calculation is completed, the diagnostic server 32 generates a detailed diagnostic result containing information such as whether the defect exists, the defect type, and the defect location. Then, it packages this result into a response message and returns it to the edge computing unit 31 via the network 40. This process corresponds to... Figure 4 The edge computing unit 31 receives the "Return Diagnostic Result" message. After receiving the result, it can immediately display it on the on-site display device or directly control the actuators (such as marking machines) on the production line. At the same time, the diagnostic server 32 can also store the received feature data and generated diagnostic results into a historical database for long-term quality traceability, data analysis, and future retraining and iterative optimization of machine learning models.

[0043] This embodiment achieves several beneficial effects by employing this edge-cloud collaborative distributed architecture. Firstly, since the transmitted data is refined feature data rather than raw signals, the network bandwidth requirements connecting the field and the central control room are significantly reduced. Secondly, most front-end data processing tasks are completed in the edge computing units at the field, alleviating the burden on the central server and thus improving system response speed. Thirdly, this architecture enhances the system's scalability and robustness, making it easier to add more sensors and edge computing nodes to the production line without requiring large-scale modifications to the central server.

[0044] Example 3 Another embodiment of this application provides a real-time diagnostic scheme for hidden defects in amorphous ribbons, characterized by the fusion of two physical field signals, eddy current and acoustic, which focuses on detecting mechanical defects such as surface and subsurface microcracks and pores in amorphous ribbons.

[0045] In this embodiment, the system hardware architecture can be referred to Figure 1The architecture shown is a centralized processing architecture, but the configuration of the multiphysics synchronous sensing array 20 is different. Specifically, sensor A 21 is a high-frequency eddy current probe array, while sensor B 22 is a high-sensitivity acoustic sensor.

[0046] A high-frequency eddy current probe array is arranged along the width of the amorphous ribbon 10. Each probe includes an excitation coil and a detection coil. During operation, the excitation coil generates a high-frequency alternating magnetic field, inducing eddy currents on the conductive surface of the amorphous ribbon 10. When discontinuous defects such as cracks or pores exist on the surface or subsurface of the ribbon, they obstruct the normal flow path of the eddy currents, causing distortion in the distribution of the eddy current field. This distortion, in turn, affects the impedance of the detection coil. By measuring the changes in the amplitude and phase of the impedance, the presence of defects can be detected.

[0047] High-sensitivity acoustic sensors, such as piezoelectric acoustic emission sensors, are tightly coupled to a support roller on the amorphous ribbon 10 production line, or aligned with the ribbon surface via non-contact air coupling. Their function is to "listen" to the sounds emitted from within the material. When the microcracks within the amorphous ribbon 10 expand or fracture under the tension of the production line, strain energy is instantaneously released, generating transient elastic waves, i.e., acoustic emission signals. This sensor is able to capture these extremely weak signals.

[0048] The processing unit 30 can also be configured as a comprehensive diagnostic host, but the machine learning diagnostic model deployed within it has been adjusted accordingly for the signal characteristics of this embodiment. It is understood that, considering that acoustic emission signals are typical time-series signals whose characteristics are closely related to time evolution, this embodiment preferably employs a recurrent neural network model, such as a recurrent neural network model with long short-term memory units. This type of model is adept at capturing and learning long-term dependencies in time-series data.

[0049] The diagnostic method in this embodiment also follows... Figure 2 The process is shown below.

[0050] In step S101, when the amorphous ribbon 10 passes through the sensing area, the eddy current probe array continuously scans the ribbon, and its detection coil outputs a complex impedance signal (i.e., eddy current signal) that changes with time. Simultaneously, the acoustic sensor continuously monitors the area, and once an acoustic emission event occurs, it outputs a corresponding voltage signal (i.e., acoustic signal). These two signals are synchronously acquired and sent to the processing unit 30.

[0051] In step S102, the processing unit 30 performs feature extraction on the received signal. For eddy current signals, the feature extraction module analyzes the changes in their impedance plane. The extracted feature parameters may include: the amplitude of the impedance change, the phase angle, and the changes in the real and imaginary parts of the signal. These features directly reflect the size and depth information of the defect. For acoustic signals, due to their transient and sudden characteristics, the feature extraction module first identifies valid acoustic emission "events" or "impacts," and then extracts a series of feature parameters for each event, such as: ring count (the number of times the signal exceeds a threshold), signal energy (the square integral of the signal amplitude), signal duration, peak amplitude, rise time, etc. These parameters are related to the intensity and nature of the sound source event.

[0052] In step S103, the features extracted from the eddy current signal and the acoustic signal are combined and input into a pre-trained recurrent neural network model. For example, the eddy current feature sequence and the acoustic feature sequence within a time window can be used as dual-channel inputs to the model. Through its internal recurrent structure and gating mechanism, the model can comprehensively analyze the intrinsic correlation between the two signals. For instance, if a significant anomaly in the eddy current signal corresponds in time to a high-energy acoustic emission event, the model will determine with very high confidence that there is an active crack that is expanding. Conversely, if there is only an anomaly in the eddy current signal without an acoustic emission signal, it may correspond to a stable, inactive internal pore or inclusion.

[0053] In step S104, the recurrent neural network model outputs the final diagnostic result. This result can clearly indicate the defect type, such as "normal," "surface microcrack," "internal porosity," or "active crack." Simultaneously, combined with the position information of the eddy current probe array, the two-dimensional coordinates of the defect on the strip can be accurately located. The diagnostic result can also be used for real-time alarms and production process control.

[0054] This embodiment achieves a more reliable diagnosis of mechanical defects by complementing the two detection principles of eddy current and acoustic emission. Eddy current detection excels at high-resolution localization and dimensional assessment of static defects, while acoustic emission detection is highly sensitive to dynamic, evolving defect behavior. The combination of the two not only improves the detection rate of minute mechanical damage defects but also provides valuable information about defect activity, which is often difficult to achieve with any single detection method.

[0055] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for real-time diagnosis of hidden defects in amorphous ribbons, characterized in that, include: At least two physical signals of amorphous ribbon are simultaneously acquired using a multi-physics field synchronous sensing array. Extract multiple pre-selected types of feature parameters from the at least two physical signals; The feature parameters are input into a pre-trained machine learning diagnostic model, and the machine learning diagnostic model outputs a diagnostic result containing defect information of the amorphous ribbon. The defect information includes at least one of the following: whether the defect exists, defect location information, and defect type information.

2. The method according to claim 1, characterized in that, The at least two physical signals include magnetic field signals and thermal field signals.

3. The method according to claim 1, characterized in that, The at least two physical signals include eddy current signals and acoustic signals.

4. The method according to claim 1, characterized in that, The machine learning diagnostic model is a convolutional neural network model or a recurrent neural network model.

5. A real-time diagnostic system for hidden defects in amorphous ribbon, characterized in that, include: A multi-physics synchronous sensing array, configured for synchronous acquisition of at least two physical signals of amorphous ribbon; as well as A processing unit is connected to the multiphysics synchronous sensing array, and the processing unit stores a pre-trained machine learning diagnostic model. The processing unit is configured as follows: Extract multiple pre-selected types of feature parameters from the at least two physical signals; and The feature parameters are input into the machine learning diagnostic model to generate a diagnostic result containing defect information of the amorphous ribbon. The defect information includes at least one of the following: whether the defect exists, defect location information, and defect type information.

6. The system according to claim 5, characterized in that, The processing unit is a single integrated diagnostic host.

7. The system according to claim 5, characterized in that, The processing unit includes an edge computing unit and a diagnostic server.

8. The system according to claim 7, characterized in that: The edge computing unit is configured to: extract the feature parameters from the at least two physical signals; and The diagnostic server is configured to receive the feature parameters extracted by the edge computing unit and run the machine learning diagnostic model to generate the diagnostic results.

9. The system according to claim 5, characterized in that, The multiphysics synchronous sensing array includes a Hall sensor array and an infrared thermal imager.

10. The system according to claim 5, characterized in that, The multiphysics synchronous sensing array includes an eddy current probe array and an acoustic sensor.

Images