Method for testing and manufacturing watch hairsprings

Abstract

1. A method for testing a hairspring or a hairspring blank configured to form a hairspring, the hairspring having to have at least one predetermined expected resonant frequency, the test method comprising the steps of: a. applying a vibrational excitation to the hairspring or hairspring blank, the vibration excitation varying over time to cover a predetermined frequency range; b. identifying at least one characteristic of a resonant frequency of the hair spring or hair spring blank, such as a resonant peak, during or in response to vibrational excitation over a predetermined frequency range; c. subjecting the resonant frequency characteristic determined in step b to a prediction machine to determine whether the hair spring or hair spring blank is affected by a defect.

Description

[Technical Field]

[0001] The present invention relates to the field of testing and manufacturing of timepiece components, and more particularly to a method for testing and manufacturing balance springs, also known as oscillators, for timepieces. [Background technology]

[0002] The movement of a mechanical watch is regulated by a mechanical regulator comprising an oscillator, in other words an elastically deformable component, the oscillation of which determines the operation of the watch. Many watches include a regulator with a hairspring as oscillator, for example, mounted on the axis of the balance wheel and set to oscillation by the escapement. The natural frequency of the balance wheel and hairspring pair allows the watch to be regulated and depends in particular on the stiffness of the hairspring.

[0003] More specifically, the frequency f of the speed regulator formed by a hairspring of stiffness R coupled to a balance wheel of inertia I is given by the following equation:

number

[0004] The stiffness of the hairspring also defines its natural vibration characteristics, such as its natural frequency and resonant frequency. In this application, the natural frequency of an elastic system (oscillator alone or oscillator-balance wheel pair) is the frequency at which this system oscillates when operating freely, in other words, in the absence of an exciting force. Furthermore, the resonant frequency of an elastic system subjected to an exciting force is the frequency at which a local maximum can be measured in the amplitude of displacement for a given point of the elastic system. In other words, when an elastic system is excited by an excitation source with a time-varying frequency, the displacement amplitude, at any point that does not correspond to an oscillation node, follows an upward slope before this resonant frequency and a downward slope after it. Typically, during such tests, a record of the displacement amplitude as a function of the exciting frequency has a peak or resonant peak in the amplitude of displacement that is associated with or characterizes the resonant frequency.

[0005] The stiffness of a hairspring oscillator typically depends on the properties of the material from which it is made, as well as its dimensions, in particular the thickness (or width) of its coil along its bar. More specifically, the stiffness is given by the following formula:

number

number

[0006] The natural frequency of a regulating mechanism formed by a hairspring of stiffness R coupled to a balance wheel of inertia I is, in particular, proportional to the square root of the hairspring's stiffness. The main specification of the hairspring is its stiffness, which must be within a well-defined interval so that it can be paired with the balance wheel, which forms the inertia element of the oscillator. This pairing is essential for accurately regulating the frequency of a mechanical oscillator.

[0007] To ensure the watch operates reliably, it is extremely important that the oscillator's characteristics are as stable as possible. The importance of magnetic fields in the modern environment has, over the years, led watchmakers to use silicon balance springs, which are less sensitive to magnetic disturbances than metal balance springs.

[0008] Highly advantageously, microfabrication techniques allow the production of hundreds of silicon balance springs on a single wafer. In particular, it is known to produce multiple silicon oscillators with extremely high precision by using photolithography and machining / etching methods on silicon wafers. While methods for producing these mechanical oscillators typically use monocrystalline silicon wafers, wafers made of other materials can also be used, such as polycrystalline or amorphous silicon, other semiconductor materials, glass, ceramics, carbon, carbon nanotubes, or composites containing these materials. Meanwhile, monocrystalline silicon belongs to the cubic crystal class m3m, in which the coefficient of thermal expansion (alpha) is isotropic.

[0009] Silicon has a highly negative value for its first thermoelastic coefficient, as a result of which the stiffness of an oscillator made of silicon, and therefore its natural frequency, varies greatly with temperature. To at least partially compensate for this drawback, documents EP 1 422 436, EP 2 215 531 and WO 2016128694 describe hairspring-type mechanical oscillators that are generated from a core (or, in the case of WO 2016128694, two cores) made of monocrystalline silicon, so that the temperature variation in Young's modulus is compensated by a layer of amorphous silicon oxide (SiO 2 ) surrounding the core (or cores), the latter being one of the few materials with a positive thermoelastic coefficient.

[0010] When hairsprings are produced from silicon or another material by collective manufacturing on a wafer, the final functional yield is obtained by dividing the number of hairsprings whose stiffness corresponds to the pairing interval by the total number of hairsprings on the wafer.

[0011] However, the micromachining processes used in the manufacture of hair springs on wafers, more specifically etching, typically result in significant variance in the dimensions of hair springs on the same wafer, and therefore in their stiffness, even though the etching pattern is the same for each hair spring. The measured variance in stiffness typically follows a Gaussian distribution. Therefore, to optimize manufacturing yields, it is also considered to center the mean of the Gaussian distribution around the nominal stiffness value and to reduce the standard deviation of that Gaussian distribution.

[0012] Furthermore, the dispersion of stiffness is even greater between hair springs of two wafers etched at different times according to the same method specifications. This phenomenon is shown in Figure 1, which illustrates the stiffness dispersion curves Rd1, Rd2, and Rd3 of hair springs on three different wafers. In general, for each wafer, the distribution of stiffness R (with respect to the number N of hair springs with this stiffness) follows a normal or Gaussian law, with each dispersion curve centered around its respective mean value Rm1, Rm2, and Rm3.

[0013] Documents WO 2015113973 and EP 3181938 propose to solve this problem by forming a hairspring with dimensions greater than those necessary to obtain a hairspring with a predetermined stiffness, measuring the stiffness of this hairspring formed by coupling it with a balance wheel with a predetermined inertia, calculating the thickness of material to be removed to obtain the dimensions necessary to obtain a hairspring with the predetermined stiffness, and removing this thickness from the hairspring.Similarly, document EP 3181939 proposes to solve this problem by forming a hairspring with dimensions less than those necessary to obtain a hairspring with a predetermined stiffness, determining the stiffness of this hairspring formed by coupling it with a balance wheel with a predetermined inertia, calculating the thickness of material to be added to obtain the dimensions necessary to obtain a hairspring with the predetermined stiffness, and adding this thickness of material to the hairspring.

[0014] In this way, as illustrated in FIG. 2, regardless of the average stiffness Rm1, Rm2, etc. on a given wafer, the stiffness distribution curves Rd1, Rd2, etc. can be re-centered about a nominal stiffness value, Rnom.

[0015] In contrast, the manufacture of hairsprings on wafers can result in defects across several hairsprings, which can be observed to lead to large variations in oscillation behavior with respect to the normally expected behavior. In particular, it can occur that the coils remain in contact or are joined to one another. Bridging phenomena can also be observed between coils due to impurities, or parts of the hairspring can be observed to remain in contact with other parts of the wafer. Clearly, these phenomena can be considered manufacturing defects that strongly affect the oscillation behavior and make one or more hairsprings on the same wafer unsuitable for use in a watch mechanism. Furthermore, it should be noted that these defects are not easy to correct and usually lead to the part in question having to be discarded. In particular, the micro-corrections mentioned above (creating oxide layers, adding or removing material across the entire part) cannot be used to correct such point-joint or bridge defects between the part or its coils and each other, or with the wafer supporting the part.

[0016] The object of the present invention is to propose a method that does not have the above-mentioned drawbacks and that allows a faster production flow and / or less risk of contamination and / or more sampling and / or easier or faster detection of hairsprings in wafers with defects that are prohibited for use in a watch mechanism. Summary of the Invention

[0017] More specifically, a first aspect of the present invention relates to a method for testing a hairspring or a hairspring blank adapted to form a hairspring, the hairspring having to have at least one predetermined expected resonant frequency, the testing method comprising the steps of: a. applying a vibrational excitation to the hairspring or hairspring blank that varies over time to cover a predetermined frequency range; b. identifying at least one characteristic of the resonant frequency of the hairspring or hairspring blank, such as a resonant peak, during or in response to vibrational excitation over a predetermined frequency range; c. subjecting the resonant frequency characteristics determined in step b to a prediction machine to determine whether the hair spring or hair spring blank is affected by a defect.

[0018] The method according to the above embodiment includes a step consisting of identifying characteristics indicative of significant defects affecting the balance spring or blank in question. For example, if the divergence identified in the spectrum of vibration frequencies obtained is greater than a predetermined threshold, the balance spring or blank in question is determined to be usable in a watch mechanism. Therefore, using a simple measurement of the vibration spectrum obtained in response to a vibration excitation, the method can detect whether a defect significantly affects the balance spring or balance spring blank to the extent that it renders the part unusable and unreworkable. (Performed on a single or "bare" or "raw" finished or blank part, i.e., without a finish surface treatment and / or not yet assembled, for example.) Such a measurement allows for identifying defects on the part without a visual analysis or a single measurement, even before removing the part from the wafer and attaching it to the vibration mechanism, saving time and resources. In other words, while still on the wafer stage, the balance spring or blank can be categorized, for example, into a first category of usable parts or parts for possible rework and a second category of defective parts to be discarded.

[0019] The method according to the above embodiment includes a step of vibrationally exciting the hairspring or hairspring blank, followed by a step of identifying the resonant frequency to estimate, by prediction, whether a defect affects the hairspring or blank. There is no assembly of a balance wheel or another component, saving time. Furthermore, measurements are performed only on the hairspring or blank, which limits the possibility of errors and contamination induced by other components or component assemblies. Measurement accuracy is improved because there are fewer sources of variation due to other components or contamination. In other words, only the hairspring or hairspring blank is tested. Vibration excitation is applied to the part or single blank without coupling to any balance wheel, weight, or oscillating system. The method allows for testing of a single free part, offering advantages in terms of increased productivity (no assembly with an oscillating system), increased quality (no contamination of the part), and increased precision (no errors linked to other components of the oscillating system).

[0020] According to one embodiment, the defect is: one or more coils bonded or bridged to adjacent coils or to the remainder of the wafer, e.g., an insulating support; localized or non-localized porosity of the material or oxide, The interface between the silicon core and the oxide layer, which has voids, delaminations, irregularities, etc. Irregular or discontinuous thickness of oxides, etc. defects or localized shortages in the material, such as penetration defects, For example, excess material associated with masking defects, Inhomogeneity of materials (silicon, silicon oxide), Flatness defects (smoothing) of the transverse slices of the bars that form the hairspring, Vertical imperfections in the transverse slices of the bars forming the hairspring (tapered or undercut faces), It may be a coil that is deformed, wavy, or off-center compared to its theoretical position. It should be noted that all the above defects are generally not correctable and that in an industrial context where they must be useful, no subsequent processing or correction can be envisaged. Consequently, the detection of such defects leads to the discarding of the identified part. In particular, the above defects cannot be compensated for, masked or corrected at the oscillator level, for example in the hairspring stud, by adjusting the installation length of the hairspring in the oscillator. It should be noted that these defects are not intentionally obtained and therefore cannot be compared or assimilated, and do not include, for example, manufacturing a part (e.g., a silicon hairspring) with intentionally higher or lower overall and / or uniform dimensions to allow for overall modification of stiffness and / or subsequent oxidation to obtain a part with a target stiffness and / or thermal compensation of the part. In other words, it should be noted that these defects are not and do not encompass rigidity defects of the hairspring or hairspring blank, especially if such defects are caused by a part that is intentionally too thick or too thin, so that the part (or all parts on the same silicon wafer) can be corrected, for example, as a function of a specific measurement of rigidity.

[0021] According to one embodiment, a vibrational excitation is applied to a hairspring or hairspring blank having a free end (typically the central collet). From a mechanical point of view, the vibrational excitation can be roughly considered as being applied to a mass (located at the center of gravity of the hairspring) connected by a spring (the elastic part of the hairspring) to a reference (a gripping pliers for a single hairspring, or a remainder of a substrate or wafer, e.g. made of silicon, for an undetached blank). The vibrational excitation sets the suspended mass in motion.

[0022] Also, note that if it is determined that a tested part must be separated or discarded due to a defect, this can be done on a single part without disassembling anything.

[0023] The method according to the above embodiment therefore makes it possible to test the hairspring blank during production, limiting the risk of contamination or mounting errors. Defects can be identified at an advanced stage of production. The method according to the above embodiment also makes it possible to test the finished hairspring, for example to perform a final conformity check.

[0024] Naturally, the frequency range of the spectrum obtained depends not only on the vibration excitation source but also on the sensor of the measuring instrument used. The frequency range is therefore linked to both the frequency range of the excitation and the frequency range to which the instrument (such as a vibrometer) for measuring the amplitude of the oscillations is sensitive. However, the frequency range of the excitation is selected so as to include at least one resonant frequency of the hairspring or blank to be tested.

[0025] The predetermined resonant frequency that the finished balance spring should have can be a target natural frequency or resonant frequency, or a target natural frequency range, or a target resonant frequency range, defined by a tolerance around a target value.

[0026] In the above methods, the resonant frequency characteristic is a characteristic of the oscillatory response measured over a predetermined frequency range that includes at least one resonant frequency. Such a characteristic is typically determined after processing the raw measurement signals (e.g., measurements of the amplitude, velocity, or acceleration of the displacement of a particular point on the hairspring or hairspring blank), which processing may include, for example, a Fourier transform to identify the resonant peak and thus the resonant frequency.

[0027] In particular, the identified characteristics may be the resonant frequency value, the width of the resonant peak, the amplitude of the resonant peak, the presence or absence of the resonant peak, the shape of the resonant peak (such as closely spaced peaks with several slope reversals). - the characteristics obtained from this spectrum of vibration frequencies of the hairspring or blank obtained in response to vibration excitation (resonance frequency, width of the resonance peak, amplitude of the resonance peak, presence or absence of a resonance peak, shape of the resonance peak, etc.); -based on a comparison between the same characteristics (resonant frequency, width of resonant peak, amplitude of resonant peak, presence or absence of resonant peak, shape of resonant peak, etc.) obtained from a reference spectrum, which may be a spectrum constructed from calibrated components, test averages and / or simulations or numerical calculations.

[0028] According to one embodiment, step c. comprises at least one step of comparing the spectrum of vibration frequencies of the hairspring or hairspring blank obtained in response to step a. with a reference spectrum, so as to determine whether the characteristic identified in step b. is an anomalous characteristic that deviates from the same characteristic of a predetermined expected resonance by a predetermined difference.

[0029] According to the above-described embodiment, by measuring the vibration response of the hair spring or the hair spring blank in response to a vibration excitation process of the hair spring or the hair spring blank and detecting anomalous characteristics of the resonant frequencies by comparison with the spectrum of vibration frequencies obtained in the reference spectrum, it is possible to estimate by prediction whether a defect affects the hair spring or the blank.

[0030] According to one embodiment, in step a) the frequency range is applied to a plurality of hairsprings or hairspring blanks simultaneously. The vibration excitation can typically be imposed on a wafer supporting, for example, several hundred hairspring blanks still attached to the wafer, thereby improving rapidity.

[0031] According to one embodiment, the frequency range is: -centered on a predetermined resonant frequency, and - Predetermined to encompass at least one frequency range having a range of at least 30% of the predetermined resonant frequency, in other words, ±15% of the predetermined resonant frequency. For example, if the predetermined resonant frequency is 1 kHz, the frequency range would be 850 Hz to 1150 Hz.

[0032] According to one embodiment, the hairspring has at least two predetermined expected resonant frequencies and the frequency range is predetermined to cover the at least two predetermined expected resonant frequencies. By covering or scanning a wide range of frequencies, multiple resonant peaks (or resonant frequencies) can be measured, which can provide greater accuracy.

[0033] According to one embodiment, step a. comprises the use of a source such as a piezoelectric source making it possible to induce or impose an acoustic excitation on a slice of the wafer supporting the hairspring blank, or preferably above or even below the hairspring or hairspring blank to be excited in particular.

[0034] According to one embodiment, the acoustic source can be coupled to an excitation cone selected to excite at least one hairspring or hairspring blank. Preferably, if the wafer supports multiple hairspring blanks, the acoustic source can be coupled to an excitation cone selected to excite at least a portion, preferably all, of the hairspring blanks.

[0035] According to one embodiment, the acoustic source is configured to generate a predetermined frequency range: - have an amplitude sufficient to generate oscillations of the balance spring or balance spring blank with an amplitude sufficient to be detected by means for measuring the displacement amplitude, velocity or acceleration of at least one point of the balance spring or balance spring blank; and / or - can be selected and / or adjusted to generate a vibration excitation that varies over time to cover a range of frequencies over a duration sufficient to estimate the vibration spectrum of the hairspring or hairspring blank.

[0036] According to one embodiment, the defect affecting the hairspring or the hairspring blank is a defect that alters the expected resonance mode, Step c. involves searching for an anomalous resonance peak between two expected normally adjacent or consecutive resonance peaks. The applicant has realized that defects in the joined coils, shortages in material, impurities linking the coils together, or other defects can cause the hairspring or blank to have resonance modes not observed in a compliant hairspring. This method is therefore intended to identify the presence of resonance peaks that are not normally present in the expected frequency spectrum. If a defect causes the emergence of a new resonance mode, the method makes it possible to detect the defect even if the other resonance peaks are correct.

[0037] According to one embodiment, the defect affecting the hairspring or the hairspring blank is a damping or amplification defect of the expected resonant mode, Step c. involves searching for anomalous resonant peaks having amplitudes that differ by at least 30% from the expected amplitude of the expected resonant peak. Applicant has realized that defects in the joined coils, lack of material, or impurities connecting the coils together can cause the hairspring or blank to have resonant modes that differ significantly from those that would be observed in a compliant hairspring. Therefore, the method is interested in quantifying the difference between the amplitude of a resonant peak in the obtained frequency spectrum and the amplitude of the same resonant peak in a reference spectrum. Alternatively, the area under the obtained resonant peak can be calculated and compared with the area under the peak in the reference spectrum; if the areas differ by, for example, 30%, it is possible to declare a defect.

[0038] According to one embodiment, the method includes a step d., which consists of categorizing the defects identified in step c. For example, if the abnormal characteristic is the presence of an unexpected resonant peak, or if the abnormal characteristic is a strong attenuation of the expected resonant peak, the prediction machine can identify one defect and distinguish it from another to categorize the defective part. The defect category can be, for example, a bonded coil, a material defect, or otherwise an unintended contact between the blank and the rest of the wafer.

[0039] According to one embodiment, step b) is based on measuring over time the amplitude or velocity or acceleration of the displacement of at least one point of the hairspring or hairspring blank, preferably at least partly carried out during step a).

[0040] According to one embodiment, step b comprises: - determining the resonant frequency of the hairspring or hairspring blank as a function of the operational or modal deformation of at least one point of the hairspring or hairspring blank, the operational or modal deformation typically being defined by the displacement amplitude or velocity, or acceleration and oscillation direction (outside or within a particular plane) as a function of excitation frequency.

[0041] According to one embodiment, the hairspring or hairspring blank is contained within a base plane, and step b comprises: - step b' measuring the amplitude, velocity or acceleration of the displacement of at least one point of the hairspring or hairspring blank in a direction perpendicular to the base plane, and / or - step b'' of measuring the amplitude, velocity or acceleration of the displacement of at least one point of the hairspring or hairspring blank in a direction contained in the base plane.

[0042] Measuring displacement or velocity in several directions allows for better identification of peaks and resonant frequencies.

[0043] According to one embodiment, - for a first predetermined resonance frequency, only step b' of measuring the displacement or velocity of at least one point of the hairspring or hairspring blank in a direction perpendicular to the base plane is carried out, and / or - for a second predetermined resonance frequency, only step b'' is carried out of measuring the displacement or velocity of at least one point of the hairspring or hairspring blank in a direction contained in the base plane.

[0044] Depending on the resonant frequency, a choice can be made to measure in one direction or another to measure the largest possible displacement or velocity in order to minimize measurement errors. More specifically, as a function of the geometry of the hairspring or hairspring blank, the mode of vibration (typically the direction of vibration) in response to a vibration excitation can change.

[0045] According to one embodiment, step b. - determining the resonance peak of the hairspring or hairspring blank as a function of the displacement amplitude or velocity of at least one point of the hairspring or hairspring blank.

[0046] According to one embodiment, the characteristics of the resonant frequency are determined based on the width of the resonant peak at the mid-height of its maximum, this processing method making it possible to limit the calculation errors that may be made if based solely on determining the frequency position of the peak defined by its maximum.

[0047] According to one embodiment, the prediction machine may be a device that makes it possible to predict, and therefore pre-empt or calculate, the presence of a defect based on one or more measured resonant frequencies, without coupling the hairspring to the balance wheel and without performing any other tests alone other than vibration excitation of the component. The output obtained by the user is information about the presence / nature or conformance / non-conformance of the defect that can be displayed or transmitted to the user of the test method. The prediction machine may be provided as follows: - based on a mathematical model (e.g. a polynomial relating one or more resonant frequencies to specific parameters such as stiffness), using a neural network to receive as input values ​​or graphs obtained from the vibration measurement spectrum to give as output the presence of defects, -Use artificial intelligence to enforce the system. In other words, the predictor machine is not a tuning or self-tuning or feedback loop system for adjusting the resonant frequency in response to a comparison of the measured value with a target value during the tuning process of the oscillator.

[0048] According to one embodiment, the prediction machine performs a classification, for example carried out by a neural network, to predict whether a defect affects the hairspring or the hairspring blank.

[0049] According to one embodiment, the method comprises a preliminary step consisting of taking into account the material of the hairspring or hairspring blank and adjusting the maximum amplitude of the vibration excitation and / or the frequency range of the predetermined frequency range as a function of the material of the hairspring or hairspring blank.

[0050] According to one embodiment, the frequency range spans a frequency range of 0 Hz to 100 kHz, preferably 0 Hz to 50 kHz, more preferably 0 Hz to 40 kHz, and highly preferably 10 kHz to 35 kHz. The applicant has observed that the prediction accuracy is better for peak or resonant frequencies located in the high frequency range. More specifically, focusing on stiffness, its effect on resonant frequency is stronger in the high frequency range (e.g., 10 kHz to 35 kHz), and therefore the sensitivity and accuracy are better over this particular range.

[0051] According to one embodiment, steps a. and b. are synchronized. Such synchronization offers the possibility to detect phase shifts, or damping, or couplings, taking into account which can improve the accuracy of the prediction or make it possible to adjust or recalibrate the vibration excitation source.

[0052] According to one embodiment, if step c. determines that the defect affects the hairspring or the hairspring blank, the method comprises at least one step consisting of identifying, separating, reworking or discarding the hairspring or the hairspring blank.

[0053] A second aspect of the present invention relates to a method for manufacturing a hairspring having at least one predetermined expected resonant frequency, the method comprising: A / forming at least one hairspring or hairspring blank having dimensions within predetermined tolerances necessary to obtain a predetermined expected resonant frequency; B / Testing the hairspring or hairspring blank according to the test method of the first aspect.

[0054] According to one embodiment, the method of manufacture comprises: C / Identifying, separating, reworking or discarding the hairspring or hairspring blank formed during step A / in accordance with the defect identification of step c. of the first embodiment.

[0055] According to one embodiment, the hairspring blank is formed on a wafer along with a plurality of other hairspring blanks.

[0056] A third aspect of the present invention relates to a method for training a predictive machine implementing step c. of the testing method of the first aspect of the present invention, the method comprising: i—forming the hairspring or hairspring blank, ii - applying to each of the hairsprings or each of the hairspring blanks a vibrational excitation that varies over time to cover a predetermined frequency range; iii- determining at least one characteristic of the resonant frequency of each hairspring or each hairspring blank during or in response to the application of a predetermined frequency range; iv' - placing a plurality of hairsprings or hairspring blanks in an oscillating mechanism having a predetermined inertia so as to measure the free oscillation frequency for each hairspring or hairspring blank; and / or iv'' - modelling, in a simulation tool, a plurality of hairsprings or hairspring blanks in an oscillating mechanism with a predetermined inertia, so as to calculate the free oscillation frequency for each hairspring or hairspring blank; v - estimating the expected resonant frequency of at least one of the resonant frequencies determined in step iii and / or the free oscillation frequencies measured in step iv' and / or calculated in step iv'', vi-To the predictor machine, for each balance spring or blank: -the characteristics of the resonant frequency identified in step iii-; - providing the same characteristics of the expected resonant frequency.

[0057] This training phase can build calibrated reference data for later comparison during the prediction / generation phase involving defect search. In particular, training allows for obtaining test data of a reference part or a part that is tested / simulated in parallel, for example, to build a reference spectrum. [Brief explanation of the drawings]

[0058] Other details of the invention will emerge more clearly from reading the following description, made with reference to the accompanying drawings, in which: [Figure 1] The uncorrected stiffness dispersion curves of the hairspring on three different wafers are shown. [Figure 2] The average centering of the stiffness on one wafer around the nominal value is shown. [Figure 3A] FIG. 1 is a simplified diagram of a method for manufacturing a mechanical oscillator, here a hairspring, on a wafer. [Figure 3B] FIG. 1 is a simplified diagram of a method for manufacturing a mechanical oscillator, here a hairspring, on a wafer. [Figure 3C] FIG. 1 is a simplified diagram of a method for manufacturing a mechanical oscillator, here a hairspring, on a wafer. [Figure 3D] FIG. 1 is a simplified diagram of a method for manufacturing a mechanical oscillator, here a hairspring, on a wafer. [Figure 3E] FIG. 1 is a simplified diagram of a method for manufacturing a mechanical oscillator, here a hairspring, on a wafer. [Figure 3F] FIG. 1 is a simplified diagram of a method for manufacturing a mechanical oscillator, here a hairspring, on a wafer. [Figure 4] Represents a device that allows assessment of the torque of a hairspring. [Figure 5] 10A and 10B illustrate schematic representations of an embodiment of the evaluation of the stiffness of a hairspring by vibration analysis. [Figure 6] 10 depicts an example of a frequency applied to a silicon wafer supporting a hairspring blank to impose a vibrational excitation. [Figure 7] Figure 6 shows an example of measuring the amplitude of the displacement of a point on a hairspring blank as a function of an imposed frequency range. [Figure 8] 8 shows in detail the resonance peaks identified at particular frequencies in FIG. 7. [Figure 9] For a defect-free part, the resonance peaks measured and superimposed for a particular frequency in FIG. 8 are shown. [Figure 10] 10 depicts an example of a predictive model built from data extracted from FIG. 9 for a defect-free part. [Figure 11] 1 illustrates a first example of measurements performed on a set of parts, including non-defective and defective parts. [Figure 12] 10 shows a second example of measurements performed on a non-defective part and a defective part. DETAILED DESCRIPTION OF THE INVENTION

[0059] 3A to 3F are simplified diagrams of a method for manufacturing a mechanical oscillator 100 on a wafer 10. The oscillator is intended in particular to equip a regulating mechanism of a timepiece component, and according to this example is in the form of a silicon hairspring 100 intended to equip the balance wheel of a mechanical timepiece movement.

[0060] Wafer 10 is illustrated in FIG. 3A as an SOI (silicon on insulator) wafer and includes a substrate or "handler" 20 having a sacrificial layer 30 of silicon oxide (SiO2) and a layer of single-crystal silicon 40. By way of example, substrate 20 may have a thickness of 500 μm, sacrificial layer 30 may have a thickness of 2 μm, and silicon layer 40 may have a thickness of 120 μm. Layer 40 of single-crystal silicon may have any crystal orientation.

[0061] The lithography process is illustrated in Figures 3B and 3C. The term "lithography" refers to a series of operations that allow an image or pattern on or above wafer 10 to be transferred back. Referring to Figure 3B, in this exemplary embodiment, layer 40 is covered with a protective layer 50, for example of a polymerizable resin. This layer 50 is typically structured by a photolithography process using an ultraviolet light source and, for example, a photomask (or other type of exposure mask) or a stepper-reticle system. This lithographic structuring forms a pattern for a plurality of oscillators in layer 50, as illustrated in Figure 3C.

[0062] 3D, the pattern is then machined, in particular etched, to form a plurality of vibrators 100 in layer 40. Etching can be performed by deep reactive ion etching (DRIE). After etching, the remaining portions of protective layer 50 are subsequently removed.

[0063] 3E, the oscillator is released from the substrate 20 by locally removing the sacrificial layer 30 or by etching all or part of the silicon of the substrate or the handler 20. Smoothing of the etched surface (not shown) may also be performed before the release step, for example by a thermal oxidation step followed by a deoxidation step consisting of a wet etch using, for example, hydrofluoric acid (HF).

[0064] 3F, the coil 110 of the silicon oscillator 100 is covered with a layer 120 of silicon oxide (SiO2), typically through a thermal oxidation step, in order to produce a temperature-compensated oscillator. The formation of this layer 120, which generally has a thickness of 2 to 5 μm, also influences the final stiffness of the oscillator and must therefore be taken into account during the preceding steps in order to obtain the vibration characteristics of the hairspring that allow obtaining a specific natural frequency of the hairspring-balance wheel pair in a given watch mechanism.

[0065] As mentioned above, in the preceding stages of creating the temperature compensation layer, the various resonators formed in the wafer generally have a large variation in geometry among the resonators, and therefore a large variation in resonator stiffness, even though the steps of forming patterns and machining / etching through those patterns are the same for all resonators.

[0066] Furthermore, this dispersion in stiffness is even greater between the balance springs of two etched wafers at different times, even when the same method specifications are used.

[0067] Finally, it should be noted that more point-like manufacturing defects can occur during manufacturing. For example, material can remain between two adjacent coils during the machining process shown in FIG. 3D. In contrast to the part shown in FIG. 3E, it is possible to see a part where material residue still exists between the coils or the substrate. Material bridges can also form between two adjacent coils during oxidation, as shown in FIG. 3F. Finally, contamination can also occur due to debris or particles adhering between the two coils or between the coil and the substrate. All of these defects strongly affect vibration behavior and cannot be corrected by adding or removing material across a set of parts, as is known in the prior art.

[0068] While the above description relates to a silicon oscillator 100, glass, ceramic, carbon nanotube, or even metal oscillators are also possible. In particular, conventional steel balance springs can be tested. In this case, the metal balance spring is clamped or sampled with reference to a tool that positions the metal balance spring opposite a radiation source and a displacement measuring device.

[0069] In a known manner, measuring the stiffness of a hairspring can be carried out in a so-called static manner, in other words by determining its torque without causing the hairspring to oscillate, see for example document EP 3654111.

[0070] An alternative to the method described in this last document involves performing torque measurements using a rheometer, such as that marketed by Anton Paar. A device provided for this purpose is illustrated in FIG. 4. Advantageously, it allows the hairspring 200 to be evaluated to be placed on a mount 202 and positioned in such a way that the hairspring can be secured at its last coil by a retaining member 204. Once the last coil is secured, the mount 202 moves away from the hairspring 200, thus completely freeing the hairspring from any elastic constraint. A rheometer head 206 is then positioned facing the collet of the hairspring. It has a counter-shape to the non-circular collet, but its reduced size allows the rheometer head to engage within the collet with controllable precision without coming into contact with the hairspring. The rheometer head is then rotated in the direction of the hairspring contraction. When the head comes into contact with the collet, it drives the collet and the rheometer measures the torque exerted by the elastic return of the hairspring over a given angle. However, such measurements remain unitary measurements and involve many risks such as breakage, contamination, etc., and require long processing times.

[0071] The invention proposes in step 3E to determine the resonant frequency of a sample of resonators 100 on the wafer based on at least one characteristic and whether defects are present on the components, in which case the invention proposes to identify the components in question without disassembly or measurements on a test subassembly, according to a method more efficient than the prior art methods.

[0072] Therefore, the present invention proposes determining at least one characteristic of the resonant frequency of a sample of oscillators by vibration measurements and applying a predictive method (for example a numerical model or a classification or categorization method) in order to link the results of said vibration measurements to the identification of possible defects present.

[0073] For this purpose, the stylistic properties of the hairsprings mounted on the wafer are utilized. During the learning phase, it is possible to arrange the prediction machine by establishing, through analytical and numerical methods, a predictive model that links manufacturing defects (bonds, bridges, contamination, etc.) to certain specifically selected frequencies (natural or resonant frequencies associated with the width at the resonant peak or mid-height).

[0074] Once the learning phase is complete (the modes and excitation frequencies to be used have been determined), it is possible to move to a prediction phase in which the generated wafer of transducers is tested and a prediction machine is used to predict whether some parts will be defective by utilizing a prediction model, and if appropriate, identify defects in order to, for example, scrap the parts.

[0075] Therefore, once each oscillator has been coupled to the balance wheel of a given watch mechanism, it is possible to incorporate testing methods into the manufacturing process in order to separate defective parts from non-defective parts and parts suitable for obtaining a particular predetermined natural vibration frequency, if necessary. Vibrational Excitation

[0076] Measuring the oscillatory response allows one to deduce at least one characteristic of the resonant frequency, for example, its value. In detail, it is first necessary to apply a vibrational excitation to the wafer. Several options are offered: a. Frequency domain measurements: 1 - Use a piezoelectric source (or any other source making it possible to induce or apply an acoustic excitation) on a slice of wafer, specifically above or below the (preferably) excited hairspring blank 200, which is excited at a specific frequency f0 (continuous single frequency excitation). In this alternative, the excitation is maintained. 2. Alternatively, it is also possible to use a piezoelectric source (or any other source capable of inducing or imposing acoustic excitation) on a wafer slice, specifically above or below the (preferably) excited hairspring blank 200, exciting it with a frequency that varies over time to cover a predetermined frequency range, e.g., 0-100 kHz, preferably 0-75 kHz, preferably 0-50 kHz, preferably 5 kHz-50 kHz, and preferably 10-35 kHz. The entire frequency range can be scanned or covered in a time interval ranging from a fraction of a second to several seconds. For example, the frequency range can be scanned or covered in less than 0.5 seconds, less than 1 second, or less than 1.5 seconds. In this alternative, the excitation frequency is continuously varied. b. Measurement in the time domain: An excitation hammer (or any other source capable of inducing pulsed acoustic excitation) is used on a slice of the wafer, specifically above or below the hairspring to be excited (preferably), to provide the shortest possible acoustic pulse (multi-frequency pulse excitation). In this alternative, the excitation is temporary and not sustained.

[0077] Furthermore, the measurements can be carried out by following a specific sampling, for example according to a sampling range of 4, 2 or 1 Hz. More specifically, the resolution for processing the acquired data, for example according to a Fourier transformation, directly depends on the duration of this acquisition.

[0078] Furthermore, if the frequency range extends, for example, up to 50 kHz, a sampling frequency of the signal of at least 100 kHz can be chosen.

[0079] Generally, finally, it may also be provided to vary the direction of excitation, in other words the direction of motion imposed by the source (vibrations may be imposed in one or more axial directions, and this or these directions may be varied over time). In the case where a wafer comprising multiple oscillators is excited, the direction of vibration may be adjusted towards one or the other of the oscillators depending on the displacement amplitude measurements described below.

[0080] Finally, it may be provided to couple an acoustic source to a divergence cone directed towards the transducers to be excited, and to adjust the acoustic source in turn to emit an excitation signal having an amplitude sufficient to impose vibrational excitation of one or more transducers and to be accurately detected and measured by selected measurement equipment. Measurement of displacement amplitude, velocity or acceleration

[0081] During excitation, the amplitude and phase (relative to the excitation source) of the oscillations in the three directions X, Y (in-plane) and Z (out-of-plane) of the specifically excited hairspring are recorded via suitable measuring means. Without limitation, the following possible measuring means can be cited: -Interferometric optical method: a.3D Doppler effect (laser Doppler vibrometer), b. Holography, - strobe optical method, -High resolution first-order chromatic confocal profilometry, -Optical reflectance measurements, a. Vibration analysis by beam deflection with a multi-dial detector or camera; b. Analysis by time-correlated TCSPC analysis; -Acoustic method by Doppler ultrasonography.

[0082] 5 schematically represents a silicon wafer 25 on which a plurality of hair spring blanks 200 are formed. A vibration excitation source 400 is coupled to the wafer 25 to impose a vibration excitation. As a result, each hair spring blank 200 vibrates, and a laser vibrometer 300, here focused on a point on the right-hand hair spring blank 200, can measure the vibration amplitude of the measurement point over time. Displacements can be measured in a direction perpendicular to the plane of the wafer 25, but displacements can also be measured in one or more directions contained in the plane of the wafer 25.

[0083] Once a particular point has been investigated, the laser vibrometer 300 can be moved onto another measurement point on the hair spring blank 200 or to another hair spring blank 200 on the wafer 25. Of course, the hair spring blank 200 can alternatively be displaced relative to the laser vibrometer.

[0084] 6 depicts an exemplary vibration excitation over time. In the given example, the excitation frequency can be varied over time between 0 Hz and 50 kHz to impose a series of rising edges, each rising edge separated by a rest period without excitation. For each measurement point on the hairspring blank 200, multiple rising fronts (between 2 rising fronts and 60 rising fronts) can be imposed, each lasting, for example, 0.5 seconds to 2 seconds. Selecting the reference point to measure

[0085] With regard to the displacement amplitude measurements, during the learning phase, a step can be provided consisting of identifying points on the oscillator at which the oscillatory response is significant. More specifically, in the case of a hairspring subjected to vibrations, especially if the frequency varies over time, the oscillatory response causes nodes to appear on the hairspring, in other words, at specific points on the hairspring at which the displacement amplitude is low or zero. If the displacement measurements are carried out at points on the hairspring that are known to be nodes at one or more specific frequencies, the identification of the resonant frequency characteristics will be adversely affected.

[0086] It is therefore advantageous to provide a preliminary step of measuring the displacement at a number of predetermined points on the hairspring, for example at least 10 predetermined points, preferably at least 20 predetermined points, and very preferably at least 30 predetermined points. The predetermined points may be selected to be located on an orthonormal coordinate system XY in the plane of the hairspring.

[0087] At the end of the preliminary amplitude measurement step on predetermined points, a resonant frequency can be identified for each measurement point, and then a step of selecting reference points for measuring the displacement amplitude during excitation indicates that no nodes of these resonant frequencies exist. In other words, the identified nodes have at least one resonant frequency, a displacement amplitude of zero or less than a first threshold peak value, and these points forming the nodes are separated from the reference points considered for subsequent measurements. It can also be noted that the reference points vary as a function of the position of the hairspring blank 200 on the wafer 25.

[0088] Typically, it is contemplated that at least two reference points are selected, and preferably at least four reference points are selected. If the transducer has a radius Ra and is fixed or embedded on the wafer by its externally pinned end, preferably four reference points may be selected and positioned: Within the first zone (e.g., on the central bead) below -0.20 × Ra, or Within the second zone between -0.05 × Ra and 0.30 × Ra (for example, on the second coil starting from the bead), or Within the third zone between -0.35 × Ra and 0.65 × Ra (for example, on the centrally located coil of the balance spring), or Within the fourth zone of -0.65 × Ra to 0.85 × Ra (for example, on a coil placed three-quarters of the way along the balance spring). Therefore, the reference point is separated from the part fixed on the wafer and necessarily has a significant capacity to the oscillatory displacement, ensuring better accuracy of the displacement measurement.

[0089] Additionally, displacement can also be measured relative to a point on the body of the wafer and / or a point on the excitation source, for example, to identify or measure phase shift or vibration damping, or even vibration coupling or resonance arising from the wafer. These complementary measurements can ensure that identified peaks are in fact peaks of the hairspring alone. Displacement amplitude measurement and vibration excitation can also be synchronized.

[0090] Alternatively, it is possible to measure only the displacement / movement / vibration at specific points, preferably located on zones of the part that do not deform. In particular, this can be provided for by measuring points on the collet of the hairspring or hairspring blank. More specifically, the collet can be considered as non-deformable during vibration excitation, and all points on the collet have the same displacement / movement / vibration. As a result, small errors in the location of the measurement points on the collet have little effect on the final result. Furthermore, by selecting specific measurement points on the part, it is possible to identify and select a specific frequency range for performing stiffness predictions.

[0091] An embodiment in which multiple parts are tested in succession while still attached to a substrate or tool includes: - capturing an image of the part to be tested; - analyzing the images, for example to identify the type of each part and / or the location of each part, - selecting one or more points to be measured for each part and / or selecting a vibrational excitation spectrum to be imposed for each part and / or each selected point, For each part to be tested, it may include a step of positioning the substrate or tool supporting the part to be tested in the vibration excitation and measurement device. According to this embodiment, in the case of a wafer still carrying the hairspring blank, the excitation and measurement can be automated: - one or more images of the wafer are taken; - an automatic image analysis is carried out to know at least the XY position of each part (part type or model recognition may also be carried out), - specific pre-established measurement points are identified or selected (e.g. on the bead) as a function of the position and / or the recognized part type; specific excitation cycles can also be selected as a function of the part type or specific points; For example, using a tool with a table that can carry the wafer and move in XY, each hairspring blank is automatically and successively positioned opposite an excitation source and measuring device to be tested by aiming at the correct measurement point and applying the correct excitation specifications.

[0092] According to one embodiment, a step may be provided that consists of providing a specific orientation for the excitation direction and / or the measurement direction as a function of the measurement point selected on the part to be tested, and / or as a function of the excitation frequency, and / or as a function of the model of the part to be tested. For this purpose, the excitation direction (or the axial direction of the excitation source) may be selected to be perpendicular to the part to be tested in order to maximize the displacement perpendicular to the plane formed by the stationary part. The excitation direction (or the axial direction of the excitation source) may be selected to be inclined with respect to the part to be tested in order to maximize the displacement included in the plane formed by the stationary part. With respect to the measurement, the measurement direction (or the axial direction of the laser beam of the measuring device) may be selected to be perpendicular to the part to be tested in order to maximize the measurement accuracy of the displacement perpendicular to the plane formed by the stationary part. The measurement direction (or the axial direction of the laser beam of the measuring device) may be selected to be inclined with respect to the part to be tested in order to maximize the measurement accuracy of the displacement included in the plane formed by the stationary part.

[0093] According to embodiments in which multiple components are attached to a substrate such as a wafer, sampling can be performed by removing one or more components in order to test them in isolation, and by introducing specific excitation frequencies to be applied, and / or specific measurement points to be used, and / or specific ranges of the vibration spectrum to be considered. In other words, this pre-sampling makes it possible to test the isolated components under favorable conditions (measurement errors and interferences are limited) in order to select the best test conditions for the components that remain integrated with the substrate. Determining vibration characteristics

[0094] Depending on the range preselected for the excitation, there are several scenarios: a. Measurement in the frequency domain Alternative example that maintains 1-excitation: i. Integrating the amplitude and oscillation phase over time long enough to have good spectral resolution at the excitation frequency f0; ii. Offset the oscillation frequency by delta f to excite at frequency f + Δf and repeat integration step i; iii. Reconstruct the amplitude and oscillation phase spectrum as a function of excitation frequency (possibly with multiple peaks at multiple frequencies). 2-Alternative example with excitation whose frequency varies over time: i. Recording vibration amplitude and phase versus time over the course of a frequency scan of a frequency range; ii. repeating step i- at ​​least once, preferably at least three times; iii. Reconstruct the amplitude and oscillation phase spectrum as a function of excitation frequency (possibly with multiple peaks at multiple frequencies). b. Time domain measurements: i. Record the displacement of the coil over time in X, Y, and Z for a duration long enough to obtain a sufficiently representative signal, for example, a few seconds. ii. One can choose to record the signal to create a reference signal to be compared with other signals measured on other components. One can also choose to perform Fourier transform type processing of the signal to identify resonant frequencies in the recorded signal.

[0095] As a result, at least one resonant peak can be identified for each excited oscillator. It is proposed to determine the resonant frequency not based on the tip of the resonant peak, i.e., its maximum amplitude, but rather on the region of the curve located between 25% and 75% of the maximum amplitude value of the resonant peak, for example, its width at mid-height. More specifically, this processing method, focusing on the portion of the curve between 25% and 75% of the maximum amplitude value of the resonant peak, allows for limiting errors due to approximation calculations for reconstructing the singular point of the maximum amplitude and the tip portion of the resonant peak. The region of the curve located between 25% and 75% of the maximum amplitude value of the resonant peak has better accuracy than the portion greater than 75% (typically the peak), providing better precision regarding the determined exact resonant frequency. For example, to determine the resonant frequency associated with the peak in question, it is possible to take the midpoint of the segment connecting two points at mid-height of the resonant peak.

[0096] FIG. 7 shows an example of a vibration spectrum of a point of the hairspring blank 200 of FIG. 5, reconstructed from measurements of the displacement amplitude of the considered measurement point in response to the vibration excitation of FIG. 6, which is free of defects and ranges from 10 kHz to 15 kHz. The presence of three amplitude peaks at approximately 11 kHz, 12.3 kHz, and 13.7 kHz can be noted. Although not shown, when the vibration excitation scans the frequency range from 0 Hz to 50 kHz, it is typically possible to identify 10 to 30 amplitude peaks. Each amplitude peak corresponds to a resonant frequency, and the maximum amplitude varies strongly.

[0097] FIG. 8 details a process that can be performed on an amplitude peak of a defect-free part, e.g., an 11 kHz amplitude peak. The objective is to find the resonant frequency and assign it as accurate a value as possible. Instead of basing this process on the maximum value of the peak, applicants realized that better accuracy could be achieved by determining the length of the segment connecting the ascending and descending portions of the curve at the mid-height of the peak. The resonant frequency is typically the mid-value of this segment. However, interpolation can be performed on points near the resonant peak to improve accuracy and shift the selected point on the segment other than the midpoint, especially if the actual location of the resonant peak is shifted, for example, due to the selected sampling frequency.

[0098] FIG. 9 shows an example of an amplitude peak at approximately 10 kHz, constructed for approximately ten tested, defect-free hairspring blanks 200. It can be noted that the frequency location of the amplitude peak varies from hairspring blank to hairspring blank (approximately 9.8 kHz to 10.02 kHz), and the maximum displacement amplitude varies by a ratio of approximately 1 to 5. Because the tips of the amplitude peaks are not actually symmetrical, it may be prudent to determine the resonant frequency based on the width of the peak at mid-height. The width of the peak at mid-height can also be used to determine damping and compare this damping to a reference value.

[0099] In these tests in Figure 9, it was possible to estimate the following resonant frequencies: [Table 1] Determination of the stiffness and / or actual dimensions of the bars of the tested transducer

[0100] During the learning phase, it is necessary to provide data on the actual stiffness and / or dimensions of the bars of the tested oscillator in order to establish a predictive model that can receive as input the vibration characteristics (typically the resonant frequencies) and give as output stiffness and / or dimensional corrections. For this purpose, the natural frequency of the hairspring-balance wheel system can be virtually measured in an environment similar to that of the particular watch mechanism.

[0101] Two alternatives can be implemented: According to the first alternative, it is possible to couple a given balance wheel onto the oscillator still attached to the wafer and measure the natural frequency of oscillation of the oscillator-balance wheel pair in order to compare the natural frequency with the expected natural frequency and, inter alia, to calculate the actual stiffness or actual dimensions based on equations 1 to 3 above. According to the second alternative, the manufacture of the oscillator to be tested can be completed in order to individually attach or couple the oscillator to the balance wheel in order to now again measure the natural frequency of oscillation of the oscillator-balance wheel pair.

[0102] In the two alternatives above, an intermediate step can be included to determine the stiffness of each oscillator and then the actual dimensions of the bars of the oscillator being tested. In other words, by analyzing the free oscillations of a hairspring coupled to a reference balance wheel, it is possible to determine the natural or resonant frequency and then the stiffness or dimensions of the bars of the oscillator. In this approach, a laser aimed at the arms of the balance wheel or the hairspring carrier records the passage times of the arms of the balance wheel or the arms of the locating pin. From this, the period is then estimated, then the frequency, and finally the stiffness. The collected data is essentially a point cloud of passage times.

[0103] More specifically, to evaluate the stiffness of a hairspring on a wafer, several solutions have been proposed, such as those described in particular in M. Vermot et al.'s Trait de construction horlogere (2011), pp. 178-179. For example, a dynamic evaluation can be performed by coupling the hairspring to a reference balance wheel with known inertia. Measuring the frequency of the assembly allows for an accurate estimation of the stiffness of the hairspring. This evaluation can be performed on the wafer or by removing the hairspring from the wafer. The above-mentioned references and prior art provide details about this method.

[0104] Similarly, the stiffness can also be estimated from the measurement of the reaction torque on the balance spring using a rheometer. The required signal represents the change in torque over time as a function of amplitude. Analysis of the slope of this curve at low amplitudes (linear part) makes it possible to estimate the stiffness and therefore the dimensions of the bars of the oscillator. The dimensions of the bars of the balance spring can then be determined.

[0105] On the other hand, the natural and / or resonant frequencies and / or stiffness can be inferred by simulation for each resonator tested on the wafer, for which purpose the resonator can be reconstructed by numerical modeling, its vibration response to an imposed spectrum can be simulated by numerical calculation, and dimensional measurements can be carried out for each tested resonator in order to find the stiffness of the resonator.

[0106] High-resolution 3D X-ray tomography allows a point cloud to be extracted, giving the 3D material density of the balance spring and, using appropriate image reconstruction, a mapping of the balance spring's sections. These different types of data make it possible to estimate the dimensions of the bars and, by geometric methods, to infer the stiffness of the balance spring.

[0107] Another approach consists of analyzing the forced oscillations of the hairspring on a reference balance wheel using an escapement. As mentioned above, laser measurements (point clouds) of the balance arm's transit time allow the frequency to be measured and the stiffness to be deduced from the frequency. An alternative is possible based on acoustic acquisition (Witschi-type microphones) that record the impulses of the various operating phases of the escapement / anchor system. The measured data are either point clouds of the balance arm's transit time or the change in sound pressure level over time. These types of experimental data allow the period, then frequency, then stiffness, and finally the dimensions of the oscillator's bars to be deduced.

[0108] Returning to the tests discussed above in FIG. 9, stiffness measurements were performed by coupling each hairspring blank 200 to a reference balance wheel, allowing the following stiffnesses to be estimated: [Table 2] Establishing a predictive model

[0109] To be able to detect defects, firstly reference data must be established or constructed, such as for example a reference spectrum. During a learning phase, oscillation amplitude measurements are performed on the physical oscillator and the resonant frequency is identified. Then, in order to be able to link the resonant frequency measured on the oscillator to the stiffness and / or dimensional (thickness) corrections made, a correlation phase must be provided on which a predictive model can be built.

[0110] The above-described operations (vibration measurement, identification of the resonance peak, determination of the bandwidth at mid-height and its mean or correction value, and the stiffness and / or dimensions of the bar) make it possible to provide a database that can correlate the position of the hairspring on the wafer, the spectrum or oscillation period, or the bandwidth at mid-height and its mean or correction value with the effective stiffness and / or dimensions of the hairspring bar. As seen above, this database can be constructed from numerical simulations of a hairspring finite element model. These simulations can generate spectra or reference oscillation periods related to stiffness. This database can also be supplemented by experimental measurements by measuring the vibration spectrum, oscillation period, and position of the hairspring on the wafer and their associated stiffness. One of the advantages of this approach lies in the fact that the training database is strengthened over the course of testing. This makes it possible to have an adaptive model depending on the wafer and hairspring, which can contribute to reducing the standard deviation in stiffness on the wafer.

[0111] This database can be used to build predictive models, providing multiple solutions.

[0112] A numerical model, for example a polynomial model, can be constructed to calculate the actual thickness, dimensional correction or actual stiffness as a function of the value of the resonant frequency.

[0113] It is also possible to perform categorization by performing a k-means partition of the input data (results of vibration measurements, typically frequencies of resonant peaks) and the output data (stiffness and / or dimensions of the bars of the vibrator) and linking them together to establish a match.

[0114] It may also be provided to process the images of the resonance peaks using a neural network, such as a perceptron, to perform classification according to the stiffness or dimensions of the bar, the classes being defined by increments of values.

[0115] In summary, the learning phase includes a testing phase (excitation of the transducer with measurement of vibration characteristics to reconstruct the vibration spectrum and identify resonant frequencies). A measurement phase of the transducer's bar stiffness and / or dimensions is also performed. Once input data (resonant frequencies) and output data (bar stiffness and / or dimensions) for significant samples are available, a predictive model building phase can be performed.

[0116] Returning to the example discussed and described in relation to FIG. 9, the data collected is as follows: [Table 3] Linear regression modeling was performed on the above data for the first six lines and the following relationship could be established: R=0.0015F-10.894 where: R is 10 -7 Stiffness in N.mm. F is the resonant frequency in Hz. Thus, stiffness can be predicted and compared to the actual stiffness measured as shown in the table below, where for the first six lines this is the data used to build or train a linear regression, and for the last four lines this is just a prediction. [Table 4] A maximum error of 4.40% can be measured and Figure 10 shows the linear regression line for the values ​​from the first six lines.

[0117] It can be seen that it is advantageous to verify that the established prediction model has good sensitivity, in other words, that for two different input values, the model gives two distinct output values. The applicant observed that the sensitivity of the prediction model was not the same for all resonance peaks. In particular, with reference to the prediction equation established and represented in Figure 10, the leading coefficient is 0.0015 10 -7N.mm / Hz. On the other hand, the applicant has observed that the leading coefficient may be larger for high resonance frequencies, providing a better prediction sensitivity for predicting clear stiffness or dimensional correction values ​​even based on close resonance frequency values. In order to verify / confirm that it is preferable to take into account and select specific resonance peaks at high frequencies (e.g., above 5 kHz) during the learning phase, it is advantageous to then provide a step of comparing the prediction sensitivities p for predicting, as accurately as possible, stiffness and / or dimensional correction as a function of the measured vibration response.

[0118] On the other hand, the applicant has also observed that even when the resonant frequencies are close, the resonant modes (in particular the deformation and / or displacement modes of the oscillator) may differ significantly, which may also affect the sensitivity of the prediction of the stiffness and / or dimensional corrections. During the learning phase, it is advantageous to provide a step of comparing the sensitivity of the predictions in order to choose to subsequently consider such a resonant frequency rather than another resonant frequency in order to predict, as accurately as possible, the stiffness and / or dimensional corrections as a function of the vibration response.

[0119] In light of the above observations regarding considering the sensitivity of predictions, it may be provided that during the learning phase, the various identified resonant peaks are classified according to their predicted sensitivity to stiffness and / or dimensional corrections. It may then be provided to define an excitation frequency range (applied during the pure prediction phase) to include at least one or more peaks or resonant frequencies that give the best sensitivity. Thus, imposing variable vibration excitation over such a predetermined frequency range ensures that accurate predictions can be made for the identified resonant peaks, or for each of the identified resonant peaks that cross or reinforce each other.

[0120] Generally, the learning phase makes it possible to select either resonant peaks at high frequencies and / or resonant peaks corresponding to particular resonant modes that allow accurate and reliable values ​​to be predicted, and the frequency range is predetermined to include at least one resonant peak, preferably multiple resonant peaks, so that either a single prediction as precise as possible or multiple predictions (one per resonant peak deemed relevant) can be made, and then cross-checking, averaging or adjusting the predicted values.

[0121] For example, a final value can be calculated by predicting multiple stiffness or dimensional correction values ​​based on multiple peaks or resonant frequencies, and then performing a weighted average based on the predicted values ​​by assigning a weight to each predicted value, each weight being determined as a function of the sensitivity identified for each corresponding peak or resonant frequency.

[0122] Alternatively and preferably, it is possible to have just one model that takes all peak or resonant frequencies as input and returns stiffness or dimensional corrections, and a learning phase of the model is used to precisely calculate weightings for the input peak or resonant frequencies. Predicted Phase

[0123] Once the learning phase is over, it is possible to proceed to a prediction phase, for example during a method for testing an oscillator. Typically, the testing method may be performed on a hairspring blank that has been produced on a wafer and is still attached to this wafer to infer the stiffness and / or dimensions of the bars of a sample hairspring in order to determine whether dimensional corrections should be applied.

[0124] Once the model is trained, the testing procedure implemented may be as follows: 1) Identifying the location of the hairspring on the wafer and vibrating the spectrum or oscillation period (as described above); 2) predicting the stiffness and / or dimensions of the bars of the hairspring by applying a predictive model; 3) Determine whether a dimensional correction is necessary to reach the natural frequency or target stiffness.

[0125] During the test method it is also possible to quantify the exact correction to be applied, so that the manufacturing method can additionally include the above tests. 1) Knowing the effective stiffness of the hairspring, estimated according to the model and the target stiffness and / or target dimensions of the bar, and applying the necessary correction application. Repeat steps 1) and 2) of the test method to test the stiffness / dimensions of the hairspring and to confirm that the target values ​​have been reached within the tolerance thresholds, or repeat these steps and dimensional corrections until the stiffness / dimensions predicted by the model reach the target values. sampling

[0126] It is known that hundreds of hairsprings are produced on a wafer and that the dimensions of the bars of the produced hairsprings can vary according to the area of ​​the wafer. If a stiffness evaluation can be performed on a single hairspring, it is in fact performed on a sample of hairsprings distributed on the wafer.

[0127] Starting from the evaluations carried out, if the results obtained vary from hairspring to hairspring, a correction can be carried out in a uniform manner for the whole wafer or separately for each region, thus making it possible to reduce the standard deviation of the stiffness variance. Furthermore, if the stiffness is known for all hairsprings by applying a model, an optimal correction can be determined that makes it possible to reduce the overall variance.

[0128] It may also be envisaged to proceed to the evaluation of all balance springs of the wafer, in particular with vibration evaluation, as this can be done very quickly and may allow the method to be automated.

[0129] Although the above examples are given mainly based on the manufacture of hairsprings with a set bar dimension larger than the target bar dimension, it is also possible to provide for the production of hairsprings with an initial bar dimension smaller than the target bar dimension. The correction step consists, for example, in adding material as described in the aforementioned document EP 3 181 939.

[0130] The method, which consists in identifying the resonant frequency by imposing a vibration excitation only on the hairspring blank, makes it possible to obtain measurement data quickly, for example without having to carry out work to mount the balance wheel, while limiting measurement errors since only the hairspring blank is tested (there are no errors that can be linked to the balance wheel, such as its mass, its mounting position, etc.).

[0131] Applicant has also observed that, alternatively, the prediction step can be a method for detecting gross defects on components such as coils bonded or bridged together or on a substrate. Figure 11, for example, shows a frequency spectrum measured on a wafer blank, including a corrected portion (whose resonant peak is surrounded by a rectangle with a dash-dotted blending line and marked "OK") and inaccurate portions marked P1, P2, P3, P4, P5, and P6.

[0132] More specifically, during testing, each frequency spectrum obtained in response to a vibration excitation is connected by traceability to a specific component, and when each of the components from which the resonant peaks P1-P6 occur is inspected, it is found to have a defect. Consequently, by comparing each of these frequency spectra with the reference spectrum previously established during the learning phase, it becomes possible to observe significant deviations or divergences, revealing defects that would cause the component in question to be discarded.

[0133] In this example, the method includes searching for and identifying excessive resonant peaks relative to the expected frequency spectrum. Such excessive peaks can be detected if they exceed the expected noise level by at least 30%. In other words, the presence of a resonant peak between two normally adjacent or consecutive resonant peaks creates an abnormal characteristic of the resonant frequency, making it possible to identify a hairspring or blank with a coupling coil defect. Since such defects cannot be easily corrected, the part is identified and, as a rule, discarded.

[0134] 12 shows another example or type of abnormal characteristic relative to the reference spectrum. More specifically, among the curves having a single resonant peak (marked "OK") at the expected resonant frequency at approximately 24.1 kHz, the presence of a curve shown in bold can be seen along with a "double-peak" resonance. In this case, to detect a defect, the method must search for two peaks or three slope change points in the portion of the curve located above the mid-height or mid-point of the recorded maximum (instead of a single maximum or a single slope change for a defect-free portion).

[0135] Therefore, by searching for peaks in the reference spectrum that are not normally present, as in Figure 11, or for resonant peaks that are significantly different from the expected peaks (having some change in slope or maximum, or even a significantly different amplitude), as in Figure 12, parts with defects including bonded coils, bridges to the substrate, material inhomogeneities, residual deformations in the hairspring, microcracks, or even accidental contamination can be identified. Each resulting frequency spectrum can be assigned or linked to the part being tested with traceability, thereby allowing the part in question to be identified and either segregated or discarded.

Claims

1. 1. A test method for testing a hairspring or a hairspring blank configured to form a hairspring, the hairspring having to have at least one predetermined expected resonant frequency, the test method comprising the following steps: applying a time-varying vibrational excitation to the hairspring or hairspring blank to cover a predetermined frequency range; b) determining at least one resonant frequency characteristic of said hairspring or hairspring blank during or in response to said vibrational excitation over said predetermined frequency range; c) subjecting the resonant frequency characteristic determined in step b to a prediction machine to determine whether the hair spring or hair spring blank is affected by a defect.

2. 2. The testing method of claim 1, wherein step c. comprises at least one step of comparing the spectrum of vibration frequencies of the hairspring or hairspring blank with a reference spectrum so as to determine whether the characteristic identified in step b. is an anomalous characteristic that deviates from the same characteristic at the predetermined expected resonant frequency by a predetermined difference.

3. 3. The testing method according to claim 1, wherein the hairspring has at least two predetermined expected resonant frequencies, and the frequency range is predetermined to cover the at least two predetermined expected resonant frequencies.

4. the defect affecting the hairspring or hairspring blank is a defect that changes the expected resonance mode, 2. The testing method of claim 1, wherein step c. comprises searching for an anomalous resonant peak between two expected normally adjacent or consecutive resonant peaks.

5. the defect affecting the hairspring or hairspring blank is a damping or amplifying defect of an expected resonant mode, 2. The testing method of claim 1, wherein step c. comprises searching for anomalous resonant peaks having amplitudes that differ by at least 30% from the expected amplitude of the expected resonant peak.

6. 2. The testing method according to claim 1, wherein step b. is based on measuring the displacement amplitude or velocity or acceleration of at least one point of the hairspring or hairspring blank over time.

7. The hairspring or the hairspring blank is contained within a base plane, and step b. a step b' of measuring the displacement amplitude or the velocity or the acceleration of at least one point of the hairspring or hairspring blank in a direction perpendicular to the base plane, and / or 2. A testing method according to claim 1, comprising a step b'' of measuring the displacement amplitude or velocity or acceleration of at least one point of said hairspring or hairspring blank in a direction contained in said base plane.

8. Step b. A test method according to claim 6 or 7, comprising the step of identifying a resonance peak of the hairspring or hairspring blank as a function of the displacement amplitude or velocity of at least one point of the hairspring or hairspring blank.

9. 9. The test method of claim 8, wherein the resonant frequency is characterized based on a width of the resonant peak at a mid-height of its maximum value.

10. 2. The testing method of claim 1, wherein the prediction machine performs a classification performed to predict whether a defect affects the hairspring or the hairspring blank.

11. 2. A testing method according to claim 1, comprising a preliminary step consisting of taking into account the material of the hairspring or hairspring blank and adjusting the maximum amplitude of the vibration excitation and / or the frequency range of the predetermined frequency range as a function of the material of the hairspring or hairspring blank.

12. 2. The test method of claim 1, wherein the frequency range spans a frequency range of 0 Hz to 100 kHz.

13. 2. The testing method according to claim 1, wherein if step c. determines that a defect affects the hair spring or the hair spring blank, the method comprises at least one of the following steps: identifying, separating, reworking, or discarding the hair spring or the hair spring blank.

14. 1. A method for manufacturing a hairspring having at least one predetermined expected resonant frequency, comprising the steps of: A / forming at least one hairspring or hairspring blank having dimensions within predetermined tolerances necessary to obtain said predetermined expected resonant frequency; B / A method comprising a step consisting of testing the hairspring or the hairspring blank according to the test method of claim 1.

15. 15. The method according to claim 14, further comprising a step consisting of identifying, separating, reworking or discarding the hairspring or hairspring blank formed during step A / if step C / c. of claim 1 determines that the hairspring or hairspring blank is affected by a defect.

16. 16. The method of claim 14 or 15, wherein the hairspring blank is formed on a wafer together with a plurality of other hairspring blanks.

17. A learning method for a predictive machine for carrying out step c. of the testing method of claim 1, comprising: i—forming the hairspring or hairspring blank; ii - applying to each of said hairsprings or each of said hairspring blanks a vibrational excitation which varies over time to cover a predetermined frequency range; iii - determining at least one characteristic of the resonant frequency of each hairspring or each hairspring blank during application of said predetermined frequency range; iv' - placing a plurality of hairsprings or hairspring blanks in an oscillating mechanism having a predetermined inertia so as to measure the free oscillation frequency for each hairspring or hairspring blank; and / or iv'' - modelling, in a simulation tool, a number of hairsprings or hairspring blanks in an oscillating mechanism with a given inertia, so as to calculate the free oscillation frequency for each hairspring or hairspring blank; v - estimating at least one expected resonant frequency of said resonant frequency determined in step iii and / or of the free oscillation frequency measured in step iv' and / or calculated in step iv'', vi - said prediction machine, for each hairspring or blank: the characteristic of the resonant frequency identified in step iii), a learning method comprising the step of providing the same characteristics of said expected resonant frequency.

Images