Bimodal Mechanoluminescence Enables Multicolor Luminescent Tailing for Ultra‐Accurate Loading Speed Visualization and Detection

TL;DR

This paper introduces a new mechanoluminescent material that changes color based on sliding speed, enabling precise visualization and detection of mechanical loading.

Contribution

The novel use of bimodal mechanoluminescence in MgF2 phosphors for ultra-accurate loading speed detection and visualization.

Findings

MgF2 phosphors show blue emission under slow sliding and multicolor emission under fast sliding.

Integration with machine learning achieves 100% accuracy in real-time bearing speed detection.

A new method for evaluating ML lifetime is proposed using luminescent trailing during fast sliding.

Abstract

Mechanoluminescence (ML) is an emerging sensing technology capable of converting mechanical stimuli into light, eliminating the need for external power supplies or complex circuitry. The integration of mechanically induced visible spatial mapping from ML sensors with image analysis algorithms provides a powerful platform for dynamic mechanical performance evaluation. In this study, we developed MgF2 phosphors through a solid‐state sintering method, integrating exhibiting instantaneous blue ML and persistent orange ML without doping ions. MgF2 displays a unique speed‐dependent ML color distribution, producing a single blue emission under slow sliding and multicolor emission (blue at the contact point and orange in the trailing region) under fast sliding. The loading speed can be vividly visualized through variations in ML intensity and color distribution. Moreover, we manage to preliminarily evaluate the ML lifetime by analyzing luminescent trailing during fast sliding, thereby tackling a long‐standing challenge in ML lifetime characterization. Additionally, integrating a machine learning algorithm with ML images captured from MgF2 phosphors enables real‐time bearing speed detection with an accuracy of 100%. This work presents a novel strategy for ultra‐accurate loading speed detection and visualization via bimodal ML, underscoring its strong potential to advance ML‐based technologies for dynamic mechanical sensing.

MgF2 exhibits a speed‐dependent mechanoluminescence (ML) color distribution arising from distinct ML lifetimes, showing blue emission under slow sliding and multicolor emission under fast sliding. A promising method for evaluating ML lifetime is proposed based on the trailing effect. By synergistically integrating instantaneous and persistent ML with a convolutional neural network, this work achieves highly accurate loading speed recognition.

Figures

Bimodal ML from undoped MgF2. a) Schematic diagrams and the corresponding ML images of MgF2 under hand‐driven slow sliding and motor‐driven fast sliding: a blue ML emission is observed under slow sliding (top), while the contact point exhibits blue ML emission and the trailing region shows orange ML emission under fast sliding (bottom). b) Schematic diagrams of the trap‐controlled ML mechanism of MgF2. c) ML spectrum of MgF2.

Morphology and structure characterization of doping‐free MgF2. a) Crystal structure of MgF2. b) XRD patterns of sintered and unsintered MgF2. c) SEM image of MgF2. d) EDS mapping of MgF2. e) HR‐TEM of MgF2. f) EDS spectrum of MgF2. g) XPS spectra of MgF2.

ML performances of MgF2 under slow slipping and fast slipping. a) Schematic diagram and extracted ML images of MgF2 under hand‐driven slow slipping and motor‐driven fast slipping. b) ML spectra of MgF2 under forces of 2–10 N. c) Integrated ML intensity versus applied force of MgF2. d) Cycle testing under an applied force of 10 N. e) Extracted ML images of MgF2 under fast slipping. f) Integrated ML intensity versus applied force of MgF2 under fast slipping. g) ML intensity of MgF2 under fast slipping for 1000 cycle testing.

Sliding speed visualization. a) ML images (top) and corresponding intensity distribution at different rotational speeds. b) Separated blue, red, and green ML images at different rotational speeds. c) Variation of RGB values with rotational speed. d) Pixel‐dependent intensity variation at different rotational speeds. e) ML intensity as a function of pixel position. f) Separated blue (iii), red (iv), and green (v) channels of processed ML images at different rotational speeds. g) Pixel‐dependent intensity variation of blue, red, and green channels. h) ML images recorded at different temperatures. i) ML intensity at different temperatures.

Visual detection of rotational speed. a) Schematic diagram of visual detection of rotational speed in bearings. b) Flowchart of machine learning based on a convolutional neural network. c) Epochs versus test accuracy based on the machine learning algorithm. d) Linear discriminant analysis (LDA) clustering of different classes. e) Confusion matrix for validating the recognition.

Full text

Introduction

1

Mechanoluminescence (ML) is a fascinating physical phenomenon in which light emission is triggered by mechanical force applied to solid materials [1, 2, 3, 4, 5, 6]. ML‐based sensors transduce diverse mechanical stimuli (such as stretching, compression, bending, friction, and impact) into light intensity and optical images without the need for external power supplies or complex circuitry [7, 8, 9, 10, 11]. These self‐powered and spatially visualizable capabilities open up new avenues for advanced sensing technologies; and offer distinct advantages for dynamic mechanical analysis [12, 13, 14, 15]. The linear relationship between ML intensity and applied force enables quantitative force measurement; while its dynamic imaging capability allows real‐time visualization of stress distribution and crack propagation. Moreover, integrating ML‐based sensors with modern imaging instruments (i.e., complementary metal‐oxide‐semiconductor and charge‐coupled device cameras) and machine learning algorithms has driven notable advancements in their functional applications, particularly in intelligent sensing and structural diagnostics [16, 17, 18, 19, 20, 21]. Compared to conventional dynamic mechanical analysis technologies, i.e., digital image correlation (DIC) and finite element analysis (FEA), ML offers a cost‐effective and naked‐eye‐visible approach for assessing mechanical behavior [22, 23, 24, 25]. Thus, ML‐based sensors have the potential to complement or even replace traditional approaches, emerging as versatile platforms for dynamic performance evaluations.

Detecting loading speed involves evaluating the motion characteristics of engine components and is essential for numerous engineering applications [26, 27, 28]. ML provides a promising approach for characterizing loading speed owing to the luminescent trailing effect [29, 30]. Typically, when motor‐driven components are coated with ML materials, their surfaces emit light under frictional forces, and relaxation processes within the crystal lattice and polymer matrix lead to luminescent trailing during friction. In particular, rapid sliding leaves insufficient time for lattice structures and polymer chains to rearrange during deformation, resulting in stress concentration and energy accumulation, and consequently a longer luminescence tail [31, 32]. This luminescent trailing effect offers an intuitive and efficient strategy for real‐time monitoring of loading speed.

The lifetime of the ML material is another key factor that determines the length of the luminescent trailing, as persistent ML produces a longer tail during friction. A representative example is the persistent ML material SrAl_2_O_4_:Eu^3+^,Dy^3+^, a well‐known persistent ML material that exhibits pronounced trailing effects even under slow sliding conditions [33, 34, 35]. Combining two or more types of ML materials with distinct lifetimes enables bimodal or multimodal ML emission, exhibiting multicolor trailing lengths that vary with sliding speed. For instance, mechanical mixing of instantaneous yellow ML material Y_3_Al_5_O_12_:Ce^3+^ with persistent green ML material Ba_0.5_Sr_0.5_Si_2_O_2_N_2_:Eu^2+^ yields yellow emission at the contact point and a green luminescent trail during handwriting [36]. Additionally, incorporating RGB ML phosphors, including CaF_2_:Tm^3+^ (red), CaF_2_:Eu^3+^ (green), and CaF_2_:Tb^3+^ (blue), enables clear visualization of trailing‐length differences among colors, with these discrepancies becoming more pronounced as rotational speed increases, thereby facilitating strain rate visualization [28]. Despite the pronounced role of persistent ML in trailing and its potential for loading‐speed detection, the number of persistent ML materials with sufficient brightness remains very limited [37, 38, 39]. Moreover, most reported multicolor ML trailing phenomena are achieved through mechanical mixing of different ML materials or by co‐doping with transition‐metal and lanthanide ions, as summarized in Table S1. Therefore, integrating bimodal ML emissions with distinct lifetimes into a single, structurally stable material remains a significant challenge.

In this study, MgF_2_ phosphors were synthesized using a solid‐state sintering method, achieving the integration of persistent orange ML emissions and instantaneous blue ML emissions. The MgF_2_ film exhibits distinct ML color distributions at different loading speeds. Under low‐speed, handwriting‐induced friction, it emits a single blue ML emission; while motor‐driven high‐speed sliding generates a multicolor ML emission, featuring blue emission at the contact point and a pronounced orange luminescent trail. The ML intensity shows a linear correlation with the applied uniaxial force and demonstrates excellent repeatability and stability under both fast and slow sliding conditions. Moreover, the MgF_2_ exhibits the unique loading‐speed‐dependent ML color distribution, which originates from its bimodal ML with distinct lifetimes. On this base, we propose a feasible approach to evaluating ML lifetimes by correlating emission persistence with trailing length. Furthermore, by integrating the speed‐dependent ML characteristics of MgF_2_ with a convolutional neural network (CNN), we develop a loading speed detection system that achieves 100% recognition accuracy. This system enables real‐time visualization of diverse loading speeds and offers a promising strategy for engineering component diagnostics.

Results and Discussion

2

Sliding‐Speed‐Dependent ML Distribution

2.1

The schematic diagram of sliding‐speed‐dependent ML emission from MgF_2_ is shown in Figure 1a. Most reported ML phosphors and ML‐based devices generally exhibit a luminescent trailing effect due to relaxation behavior and intrinsic luminescence lifetimes, with the single ML emission color typically independent of sliding velocity. In contrast, MgF_2_ phosphors exhibit a distinctive, sliding‐speed‐dependent ML color distribution. During the process of hand‐driven slow sliding, the emission of blue light dominates, and a continuous luminescent path is recorded due to the extended camera exposure time (Figure 1a, top). It is noteworthy that under conditions of motor‐driven fast sliding, MgF_2_ phosphors exhibit multi‐color ML emission, characterized by blue light at the contact point and a distinct orange luminescent trail in the trailing region (Figure 1a, bottom). The vivid color mapping associated with loading speed is attributed to the multiple ML emissions with distinct lifetimes from MgF_2_, resulting in different luminescence trailing lengths.

Bimodal ML from undoped MgF2. a) Schematic diagrams and the corresponding ML images of MgF2 under hand‐driven slow sliding and motor‐driven fast sliding: a blue ML emission is observed under slow sliding (top), while the contact point exhibits blue ML emission and the trailing region shows orange ML emission under fast sliding (bottom). b) Schematic diagrams of the trap‐controlled ML mechanism of MgF2. c) ML spectrum of MgF2.

A plausible mechanism for ML involves a piezoelectric‐induced carrier de‐trapping model (Figure 1b). Typically, upon stress stimulation, the energy band of MgF_2_ tilts, facilitating the tunneling of trapped electrons from shallow donor levels to the conduction band. Subsequently, the conduction electrons then transition from the conduction band to the valence band, transferring energy to the luminescent centers via nonradiative energy transfer (NET). This process ultimately results in the emission of ML. The luminescent centers have been found to be associated with self‐trapped exciton (STE) emission related to fluorine vacancies, including F(C_2v_), M(D_2h_), M(C_2h_), and M(C_1_), as has been previously reported [40, 41, 42, 43]. As shown in Figure 1c, the ML spectra of MgF_2_ span a wide range from the ultraviolet to the visible region. They feature multiple sharp peaks in the ultraviolet‐blue range (250 − 450 nm) and a broad peak in the orange region (500 − 700 nm), corresponding to the blue emission at the contact point and the long orange luminescent trail, respectively.

In the trap‐controlled piezoelectric‐induced carrier de‐trapping model, both the piezoelectricity of the host material and the characteristics of the trap levels are critical for achieving high‐intensity emission. The piezoelectricity of MgF_2_ is characterized using piezoresponse force microscopy (PFM). Clear piezoelectric hysteresis and butterfly loops are observed under bias voltages ranging from −5 V to +5 V, confirming the existence of a piezoelectric response and polarization switching behavior (Figure S1). Although MgF_2_ is intrinsically centrosymmetric, the introduction of vacancies locally breaks the symmetry, thereby giving rise to measurable piezoelectricity. In addition, thermoluminescence (TL) measurements revealed trap depths of MgF_2_ are 0.7 and 0.84 eV (Figure S2). The trap depth (*E_T_ *) is estimated using the empirical equation

$$[eqn]$$

where *E_T_

• is the estimated trap depth with the unit of eV, and *T_m_
• is the temperature at which the TL curve reaches a maximum with a unit of Kelvin.

Morphology and Structure Characterization

2.2

MgF_2_ phosphors were prepared using a solid‐state sintering method (details in Experimental Section). The crystal structure of MgF_2_ crystallizes in tetragonal symmetry (space group P4_2_/mnm), with lattice parameters of a = b = 4.5967 Å, c = 3.0376 Å (Figure 2a). Each Mg^2+^ ion is octahedrally coordinated by six F^−^ ions, forming [MgF_6_] octahedra that interconnect through edge‐sharing to form a 3D framework. X‐ray diffraction (XRD) confirms the formation of phase‐pure MgF_2_, with all diffraction peaks indexed to the tetragonal phase (JPDF#41‐1443) (Figure 2b). Compared with the as‐prepared MgF_2_ powders, the sintered MgF_2_ samples exhibit sharper diffraction peaks, indicating enhanced crystallinity after high‐temperature solid‐state sintering. Such a stable crystal field environment facilitates ML emission.

Morphology and structure characterization of doping‐free MgF2. a) Crystal structure of MgF2. b) XRD patterns of sintered and unsintered MgF2. c) SEM image of MgF2. d) EDS mapping of MgF2. e) HR‐TEM of MgF2. f) EDS spectrum of MgF2. g) XPS spectra of MgF2.

After grinding, refined white MgF_2_ powders are obtained (Figure S3a). A high‐strength and robust epoxy resin is selected as the polymer matrix, and the resulting MgF_2_/epoxy composites exhibit good transparency (Figure S3b). Scanning electron microscopy (SEM) reveals that the MgF_2_ particles have irregular shapes with diameters ranging from 10 to 20 µm (Figure 2c). Energy‐dispersive X‐ray spectroscopy (EDS) mapping confirms the uniform distribution of the Mg, F, and O elements (Figure 2d). High‐resolution transmission electron microscopy (HRTEM) identifies clear lattice fringes that correspond to the crystallographic planes of (1¯01) (Figure 2e). EDS elemental analysis of MgF_2_ proves the presence of Mg and F elements, along with a minor amount of oxygen, with an atomic ratio of 2.7% (Figure 2f). Wide‐scan X‐ray photoelectron spectroscopy (XPS) spectrum also indicates the presence of a small amount of oxygen in the sample (Figure 2g). During the high‐temperature solid‐state sintering process, a small amount of oxygen atoms is introduced into the MgF_2_ sample. The oxygen exists in the form of adsorbed oxygen, interstitial oxygen, lattice oxygen, and O‐F_x_. Upon sintering, the concentration of O‐F_x_ decreases significantly, whereas the concentration of vacancies increases (Figure S4). The involvement of oxygen atoms promotes the formation of fluorine defect clusters within the MgF_2_ host, including F(C_2v_), M(D_2h_), M(C_2h_), and M(C_1_). These defect clusters may serve as plausible luminescence centers responsible for ML emission. Specifically, ML emission in the UV‐blue region is attributed to F(C_2v_), M(D_2h_), and M(C_2h_); while the orange emission band may be associated with the M(C_1_) [40, 41, 42, 43].

ML Performances

2.3

The ML images of MgF_2_ under slow sliding with diverse uniaxial forces (ranging from 2 to 10 N) are shown in Figure 3a. With increasing applied force, the ML brightness is enhanced; while the dominant emission color remains blue without any noticeable hue change. ML spectra of MgF_2_ show broad emission ranging from the ultraviolet to the visible region (Figure 3b). A slight redshift in the UV‐blue range is observed with increasing force, which is attributed to the modulation of the local crystal field under external mechanical stress. The integrated ML intensity exhibits a linear dependence on the applied uniaxial force with small errors (y = 9063x – 4736, *R^2^

• = 0.991) (Figure 3c). Moreover, cyclic tests confirm excellent self‐recovery and repeatability of ML emission (Figure 3d). Notably, ML cyclic tests are conducted without pre‐irradiation using external light sources, i.e., ultraviolet light or X‐ray irradiation). These excellent ML performances can meet more demands in practical applications.

ML performances of MgF2 under slow slipping and fast slipping. a) Schematic diagram and extracted ML images of MgF2 under hand‐driven slow slipping and motor‐driven fast slipping. b) ML spectra of MgF2 under forces of 2–10 N. c) Integrated ML intensity versus applied force of MgF2. d) Cycle testing under an applied force of 10 N. e) Extracted ML images of MgF2 under fast slipping. f) Integrated ML intensity versus applied force of MgF2 under fast slipping. g) ML intensity of MgF2 under fast slipping for 1000 cycle testing.

The multi‐color ML images are observed under motor‐driven fast sliding. With an increase in uniaxial force, the emission intensity at the contact point is enhanced, and the luminescent trail becomes longer (Figure 3e). Similar to slow sliding, the integrated ML intensity under fast sliding also shows an approximately linear relationship with the applied force, with minor deviations (y = 24610x – 27787, *R^2^

• = 0.987) (Figure 3f). Subsequent to undergoing 1,000 cycles of testing, the ML intensity exhibits a retention of over 95% of its initial value (Figure 3g). The inset illustrates the ML image after 1000 cycle tests, with minimal alterations in emission intensity. These results indicate that MgF_2_ exhibits remarkable cyclic stability, even under conditions of high loading speed friction.

Sliding Speed Visualization

2.4

ML emission intensity and luminescent trailing distribution are also influenced by the motor's loading speed. At a constant uniaxial force of 10 N, ML images of MgF_2_ are recorded at loading speeds of 150, 300, 600, 900, 1200, and 1800 rpm, corresponding to approximate linear velocities of 0.196, 0.393, 0.785, 1.178, 1.570, and 2.355 m/s, respectively (Figure 4a). The intensity distribution maps exhibit that the emission is most intense at the contact point and gradually decreases along the trailing path. Both the ML intensity and the trailing length increase with loading speed. At a speed of 1800 rpm, the elongated trailing paths connect end to end, forming a continuous luminescent ring. These results suggest that ML provides a feasible strategy for assessing loading speed.

Sliding speed visualization. a) ML images (top) and corresponding intensity distribution at different rotational speeds. b) Separated blue, red, and green ML images at different rotational speeds. c) Variation of RGB values with rotational speed. d) Pixel‐dependent intensity variation at different rotational speeds. e) ML intensity as a function of pixel position. f) Separated blue (iii), red (iv), and green (v) channels of processed ML images at different rotational speeds. g) Pixel‐dependent intensity variation of blue, red, and green channels. h) ML images recorded at different temperatures. i) ML intensity at different temperatures.

Vivid color changes in the ML images become apparent with increasing loading speed. These variations in color are more easily captured by the human eye or a camera than changes in light intensity. The original images are separated into blue, red, and green channels, as shown in Figure 4b. As the rotation speed increases, the trailing length in the red channel becomes more prominent, which can be attributed to its proximity to the orange ML emission. To quantify the ML intensity of the red, green, and blue channels, integrated grayscale values were calculated for each image using the following equation

$$[eqn]$$

where i is the number of pixels, and *G_i_

• denotes the grayscale of the i‐th pixel. The integrated grayscale values of each color channel at different loading speeds are presented in Figure 4c. With increasing loading speed, the ML intensity of all three channels increases. The ML intensity and corresponding spatial images are concurrently acquired using an optical fiber and a camera, respectively. The ML intensity recorded by the optical fiber spectrometer demonstrates a strong linear correlation with the applied loading speed, thereby enabling its quantitative monitoring (Figure S5a). Similarly, the aggregate of the integrated grayscale values (*I_R_ *+*I_G_ *+*I_B_ *) exhibits a positive correlation with the loading rate, allowing for the prediction of loading rates based on these grayscale values (Figure S5b). The positive correlation between ML intensity and loading speed indicates its potential for monitoring continuous speed variations. The observed discrepancies between the two methods can be attributed to their fundamentally different measurement principles, with the camera relying on image‐based analysis and the fiber spectrometer operating through a photon‐counting mechanism. Furthermore, analyzing the spatial distribution of ML intensity enables the detection of potential interference from ambient light, thereby improving the sensor's robustness against environmental disturbances. In addition, multichannel spatial fingerprints extracted from the RGB channels provide multidimensional information, including colorimetric ratios. The ratios of integrated grayscale intensities (*I_R_ */*I_B_
• and *I_G_ */*I_B_ *) decrease with increasing loading speed (Figure S6). These results indicate that loading speed can be evaluated using colorimetric signals in addition to ML intensity across different channels.

The characterization of ML lifetimes remains a daunting challenge due to the relatively weak emission intensity and unique stimulation method. Real‐time ML intensity can be monitored using rapid mechanical stimulation combined with photon counting devices for ML lifetime evaluation. Among these, photomultiplier tubes (PMTs) are widely recognized as an effective technique for characterizing ML lifetime owing to their superior sensitivity and high temporal resolution. However, the short photon integration time (typically nanoseconds to microseconds) necessitates sufficiently bright ML emission to ensure an adequate signal‐to‐noise ratio; and strict shielding from ambient light is essential to prevent PMT overexposure and performance degradation. Here, we propose a simple and accessible approach for evaluating ML lifetimes based on the spatial distribution of ML trailing. Although its quantitative accuracy is constrained by the relaxation dynamics of the polymer matrix, the method can be further improved through standardized testing procedures and proper sample calibration, making it suitable for comparative evaluation of ML lifetimes across different materials. A detailed comparison of these two methods is provided in Table S2.

To investigate the variation of ML intensity along the trajectory, the ML images are processed using a simple algorithm. By converting polar coordinates into Cartesian coordinates, the circular trailing path can be mapped onto a linear trajectory. As shown in Figure 4d, MgF_2_ exhibits the highest blue ML emission at the contact point, followed by an orange luminescent tail. The trailing path length increases with increasing applied force. Along the trailing trajectory, the intensity decreases nonlinearly with distance, which arises from the intrinsic lifetime of ML emission (Figure 4e, and Figure S7). The luminescence decay is typically described by an exponential function:

$$[eqn]$$

where I(t) is the real‐time ML intensity, *I_0_

• is the initial intensity, and τ is the lifetime. The sliding displacement (x) can be expressed as

$$[eqn]$$

where x is displacement, v is the sliding velocity, and 𝑡 is time. Accordingly, the relationship between ML intensity and displacement is given by

$$[eqn]$$

Accordingly, the ML intensity decreases nonlinearly with increasing sliding distance. Moreover, the experimental data can be well fitted using this equation (Figure S8). The good agreement between the experimental data and the fitted results demonstrates the feasibility of assessing ML emission lifetimes based on the luminescent trailing effect. It should be noted that the extracted τ does not represent the actual lifetime. The ML trailing length is influenced by multiple factors, including the intrinsic ML emission intensity, relaxation dynamics of the polymer matrix, and friction‐induced thermal effects. Nevertheless, the relationship between ML intensity and displacement can be well described by decay‐curve fitting, demonstrating the feasibility of this approach for qualitative lifetime analysis.

To investigate the emission lifetimes of different colors, the processed ML images are decomposed into blue, green, and red channels (Figure 4f). Along the luminescence trajectory, all three channels exhibit nonlinear decay with distinct differences (Figure 4g). The blue channel exhibits a markedly faster decay than the red and green channels. This behavior is attributed to the short lifetime of blue ML emission, whereas the red and green channels correspond to the orange trail with a considerably longer lifetime. These results demonstrate that the lifetimes of different color emissions can be distinguished through analysis of ML trailing images. Looking ahead, the application of optical filters or highly sensitive single‐band cameras may enable more precise characterization of ML lifetimes by capturing the detailed features of these trailing images.

Heat generation during friction is inevitable, especially under prolonged high‐speed sliding. To evaluate the thermal stability of ML performance, MgF_2_ is tested under a constant normal force of 10 N and a rotational speed of 1200 rpm across a temperature range of 30–150°C. Both ML intensity and the luminescent trail show only minor changes with increasing temperature, confirming the excellent thermal stability of ML emission (Figure 4h,i). These results confirm the potential of MgF_2_ for long‐term, reliable detection and visualization of loading speed of loading speed, making it a strong candidate for real‐time monitoring of critical components.

Visual Detection of Loading Speed

2.5

Bearing rotational speed is a critical parameter in many mechanical systems and automobiles. By leveraging the bimodal ML with distinct lifetimes from MgF_2_, a real‐time loading speed monitoring system is developed through the integration of a machine learning algorithm. A thin MgF_2_/epoxy coating is applied onto the motor‐driven bearing surface, and ML images are captured by a camera under uniaxial pressure (Figure 5a). The ML images are collected under dark ambient conditions to minimize interference from external light sources. These images are subsequently analyzed using a convolutional neural network (CNN) for feature extraction and classification (Figure 5b). Specifically, the ML images are resized to 128 × 128 pixels and processed through convolutional, pooling, and fully connected layers, ultimately yielding the predicted speed classifications.

Visual detection of rotational speed. a) Schematic diagram of visual detection of rotational speed in bearings. b) Flowchart of machine learning based on a convolutional neural network. c) Epochs versus test accuracy based on the machine learning algorithm. d) Linear discriminant analysis (LDA) clustering of different classes. e) Confusion matrix for validating the recognition.

The CNN is particularly well‐suited for learning complex spatial patterns, enabling accurate recognition of ML images corresponding to different loading speeds. Six predefined speed classes (V1–V6) are set at 150, 300, 600, 900, 1200, and 1800 rpm. Each class contains 200 samples for training and 300 samples for validation. As shown in Figure 5c, the model achieves 100% accuracy after only three epochs, with the training loss decreasing rapidly and then stabilizing, indicating strong generalization capability.

Furthermore, the linear discriminant analysis (LDA) clustering map constructs in the 2D feature space using hidden‐layer outputs (Figure 5d) revealed distinct, well‐separated clusters with significant inter‐class separability and intra‐class compactness. These results further confirm the effectiveness of the CNN in extracting and classifying ML features. Consequently, an ultra‐high prediction accuracy of 100% was achieved (Figure 5e). This method establishes a unique and self‐powered approach for evaluating component load speed, offering the great potential of ML materials for dynamic mechanical analysis.

Discussion

3

In this work, a doping‐free MgF_2_ ML material was developed via a solid‐state sintering method, exhibiting instantaneous blue ML emission alongside persistent orange ML emission. The MgF_2_ phosphors demonstrate outstanding ML performance, as well as excellent stability and repeatability. By synergistically combining instantaneous and persistent ML, variations in loading speed can be vividly represented through distinct ML color changes. Additionally, a simple and effective method for preliminary evaluation of ML lifetimes is proposed, based on analysis of ML images obtained during fast sliding. Integration of a convolutional neural network with MgF_2_ ML materials enables an efficient approach for accurately identifying different bearing speeds, achieving an ultrahigh accuracy of 100%. This work presents a novel and practical strategy for loading speed detection and underscores the potential of ML‐based sensing technology for dynamic mechanical analysis.

Experimental Section

4

Preparation of MgF2 Phosphors

4.1

The undoped MgF_2_ phosphors were prepared via a high‐temperature solid‐state sintering method. Typically, MgF_2_ powders (AR, Macklin) was initially ground with ethanol for 20 min, followed by drying in a vacuum oven at 80°C. The dried powders were then sintered in a tube furnace at 1200°C for 1 hour under a flowing argon atmosphere. After sintering, the samples were naturally cooled to room temperature and subsequently ground into fine powders for further characterization.

Preparation of MgF2 Composite Film

4.2

Epoxy resin was used as the polymer matrix for the MgF_2_ phosphors. The precursor solution was prepared by mixing the epoxy base agent and curing agent at a mass ratio of 3:1. The obtained solution was spin‐coated onto a polymethyl methacrylate substrate at a rotation speed of 1000 rpm, after which MgF_2_ phosphors were adhered onto the epoxy layer. MgF_2_ phosphors were incorporated into the polymer matrices at a loading of around 20 wt.%. The resulting MgF_2_/epoxy composites were cured at 80°C for 180 min, followed by post‐curing at 120°C for 60 min. Bearings with ML coatings were fabricated using the same method.

Characterization and Measurements

4.3

The morphology of the samples was characterized using scanning electron microscopy (SEM, JEOL JSM‐6700F) and transmission electron microscopy (TEM, JEOL JEM‐2100). Energy dispersive X‐ray spectroscopy (EDS) elemental mapping analyses are used to analyze the distribution of the elements on the surface. Crystal structures were analyzed via X‐ray diffraction (XRD, PANalytical X'Pert PRO) using Cu Kα1 radiation (λ = 1.5406 Å). The electronic binding energies are measured by using X‐ray photoelectron spectroscopy (XPS) (ESCALAB 250Xi, Thermo) using Al Kα radiation (hν = 1486.6 eV). Piezoresponse force microscopy (PFM) measurements were conducted using a Bruker Icon atomic force microscope. Thermoluminescence (TL) curves were measured using an LTTL‐3DS multifunction defects fluorescence spectrum (Rongfan Tech, China). Mechanoluminescence spectra were measured with a fiber optic spectrometer (Andor SR‐500i‐A). Optical photographs of ML were captured using a mobile phone with an ISO sensitivity of 12 800 and an exposure time of 3 s.

Statistical Analysis

4.4

The experiments were conducted in at least three replicates. The data were presented as mean and standard deviation of error propagated from replicates. Statistical details of experiments could be found in the figure legends where applicable.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting file: advs73691‐sup‐0001‐SuppMat.docx.