Method and device for diagnosing demagnetization fault of magnetic core of plasma source

Abstract

The invention belongs to the technical field of plasma sources, and provides a plasma source magnetic core demagnetization fault diagnosis method and device. The method comprises the following steps that the temperature change rate delta T, the ignition voltage Vloop and the gas dissociation rate lambda of a magnetic core in unit time are obtained in real time; performing normalization processing on the acquired data; inputting the data into the trained BP-MTN model, and outputting a demagnetization degree d; and estimating the demagnetization degree according to the demagnetization degree d, and processing the demagnetization degree d and the demagnetization degree d respectively. Through the diagnosis method, estimation of the magnetic core demagnetization fault of the plasma source system is realized, so that stable dissociation of gas by the plasma source is realized, the plasma concentration is ensured to be stable, and the consistency of the etching rate of a semiconductor manufacturing process is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of plasma sources, and in particular, to a method and device for diagnosing the demagnetization fault of a plasma source magnetic core. Background Art

[0002] Remote plasma source systems are widely used in semiconductor manufacturing processes. In remote plasma source systems, ferrite magnetic cores, as key components for stable plasma generation, are widely used in fields such as electromagnetic interference (EMI) filtering, energy transmission, and high-frequency circuit regulation. Their core role is to optimize the transmission efficiency of electromagnetic energy and suppress interference signals through high magnetic permeability and low loss characteristics. However, ferrite magnetic cores are prone to demagnetization under complex working conditions, resulting in attenuation or even failure of magnetic properties, which has become an important problem restricting the stability of remote plasma source systems; The essence of ferrite demagnetization is the destruction of the magnetic domain structure, manifested as a decrease in magnetic permeability, an increase in loss, and a decrease in saturation magnetic induction intensity. Research and engineering practice have found that under normal working conditions, remote plasma ferrite magnetic cores usually independently maintain their persistent magnetic fields. However, when the plasma source operates in a high-temperature environment, the increase in heat means an increase in atomic motion, which will ultimately overwhelm the alignment of magnetic domains; within the working temperature range, the magnet can withstand it for a long time without causing temporary or permanent loss of magnetic properties, but when it reaches the Curie temperature, the magnet will permanently lose all its magnetization intensity. In this case, the structure will be damaged and cannot be repaired, and it cannot be remagnetized. As the temperature of the magnetic core approaches its Curie point, demagnetization also occurs at different levels; the inductance of the magnetic core is also related to the material and structure of the magnetic core, and the magnitude of the magnetic core current is affected by the magnitude of the current passing through the magnetic core. Under specific conditions, such as when the current is too large and causes magnetic core saturation, although the current may increase, the inductance will decrease due to magnetic saturation, resulting in a weakened function of the inductor; The demagnetization problem of ferrite magnetic cores has a profound impact on system performance. For example, in an EMI filter, demagnetization will cause a decrease in the common-mode inductance value, weaken the high-frequency interference suppression ability, and thus affect the signal purity of the plasma source. In addition, demagnetization may also cause problems such as reduced energy transmission efficiency and abnormal temperature rise, exacerbating system energy consumption and thermal management difficulties. Although the anti-demagnetization ability can be partially improved through material modification (such as optimizing the composition ratio) and process improvement (such as precision sintering), the long-term stability of ferrite magnetic cores still faces challenges under extreme conditions such as high frequency, high temperature, or strong magnetic fields. Therefore, studying the demagnetization mechanism and developing targeted protection technologies have become important research directions for improving the reliability of remote plasma sources. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide a method and device for diagnosing the demagnetization fault of a plasma source magnetic core to solve or partially solve the above problems.

[0004] In a first aspect, an embodiment of the present invention provides a method for diagnosing the demagnetization fault of a plasma source magnetic core, including the following steps: A: Data acquisition: Real-time obtain the temperature change rate ΔT, ignition voltage Vloop, and gas dissociation rate λ of the magnetic core per unit time; B: Data preprocessing: Normalize the temperature change rate, ignition voltage, and gas dissociation rate; C: Input the data into the trained BP-MTN model and output the demagnetization degree d; The construction and training method of the BP-MTN model is as follows: (i) Construct a BP neural network model including an input layer, a hidden layer (using the ReLU activation function), and an output layer (using the tansig activation function); (ii) Calculate the output value d = F(ΔT, Vloop, λ) through forward propagation and calculate the error based on the mean square error function; (iii) Use the multi-dimensional Taylor network to perform Taylor expansion on the activation function, and dynamically adjust the weights and biases through backpropagation until the error converges; D: Estimate the demagnetization degree of the magnetic core according to the demagnetization degree d and process them respectively.

[0005] The beneficial effects of the above embodiments are as follows: Through this diagnostic method, the prediction of the demagnetization fault of the magnetic core of the plasma source system is realized, so as to realize the stable dissociation of the gas by the plasma source, ensure the stability of the plasma concentration, and ensure the consistency of the etching rate of the semiconductor process technology. In this diagnostic method, the BP-MTN model combines the characteristics of the control methods of both the BP neural network and the multi-dimensional Taylor network, and uses adjustable weight coefficients and biases to dynamically adjust the main control objectives; when training the BP-MTN model, by normalizing the training samples and initializing the BP neural network model, setting relevant parameters, calculating the inputs and outputs of each layer, and calculating the error, if the function converges at this time, save this BP neural network, if the function does not converge, modify the threshold and weights through the Taylor expansion of the activation function, and repeat until the function converges to construct the BP neural network model d = F(ΔT, V, λ); The BP-MTN model enables the scheduling scheme to have self-learning and adaptive capabilities through the backpropagation of the output variables in the neural network.

[0006] According to a specific implementation manner of an embodiment of the present invention, in step A: The gas dissociation rate λ is obtained by detecting the outlet spectrum with an infrared spectrometer and comparing it with a preset sample database. By comparing the infrared spectrometer with the sample database, the influence of environmental interference on the detection of the dissociation rate is eliminated, the detection accuracy is improved, and the lag of the traditional chemical detection method is avoided.

[0007] According to a specific implementation manner of an embodiment of the present invention, the normalization formula in step B is: ; Among them, Mi is the normalized input eigenvalue, Ni is the input eigenvalue before normalization, Nmax is the maximum input eigenvalue, and Nmin is the minimum input eigenvalue. The normalization process unifies multi-source heterogeneous data to the [0,1] interval, eliminates the dimension difference, improves the model convergence speed, and reduces the risk of gradient explosion.

[0008] According to a specific implementation manner of an embodiment of the present invention, in step C (iii), the output layer weight update formula of the multi-dimensional Taylor network is: ; Among them, η is the learning rate, α is the inertia coefficient, and E is the mean square error function. Introducing the inertia coefficient α suppresses the weight oscillation. Combining with the dynamic learning rate η, the model training efficiency is improved.

[0009] According to a specific implementation manner of an embodiment of the present invention, in step C (iii), the contribution degree of each input parameter to the demagnetization degree is analyzed through statistical nodes, and the weights of key parameters whose contribution degree exceeds a preset threshold are preferentially optimized. By preferentially optimizing the key parameters (such as the temperature change rate) whose contribution degree exceeds the threshold (such as ≥0.7), the sensitivity of the model to the core fault characteristics is improved, and the false alarm rate is reduced.

[0010] According to a specific implementation manner of an embodiment of the present invention, the demagnetization degree is divided into five levels: 0 (no demagnetization), 0.2 (slight), 0.5 (moderate), 0.8 (severe), and 1 (complete demagnetization) according to the output value d. The five-level quantization grading clarifies the severity of the fault, guides the grading response strategy (such as early warning, switching to a standby magnetic core), reduces the system downtime, and reduces the operation and maintenance cost.

[0011] In a second aspect, an embodiment of the present invention provides a diagnostic device for the demagnetization fault of a plasma source magnetic core, which is used to implement the foregoing diagnostic method, including: a sensor sampling unit, a dissociation rate detection unit, and a multi-dimensional Taylor network neural network module; The sensor sampling unit includes a voltage sensor for acquiring the ignition voltage Vloop and a temperature sensor for acquiring the temperature data in the reaction chamber; The dissociation rate detection unit includes an infrared spectrometer, which is used to detect the outlet spectrum and compare it with the offline sample database to obtain the gas dissociation rate λ; The multi-dimensional Taylor network neural network module runs the multi-dimensional Taylor network and the BP neural network model (BP-MTN) based on the processor, and takes the average value of the temperature change rate of the magnetic core per unit time, the ignition voltage Vloop, and the gas dissociation rate λ as the input, and the output is the demagnetization degree d of the magnetic core; the demagnetization degree is predicted according to the value of the output demagnetization degree d, and corresponding processing is performed respectively.

[0012] The embodiments of the present invention have at least the following technical effects: First, by integrating the BP neural network and the multi-dimensional Taylor network (MTN) model, high-precision dynamic diagnosis of the magnetic core demagnetization fault is achieved. Compared with the traditional single neural network method, the diagnostic accuracy is improved, and real-time online monitoring is supported.

[0013] Second, based on the normalization process and the adaptive weight adjustment mechanism of the multi-dimensional Taylor network, the model convergence speed and stability are significantly improved, the training period is shortened, and robustness is maintained under extreme working conditions such as high frequency and high temperature.

[0014] Third, through the five-level demagnetization degree classification and the analysis of the contribution degree of fault features, the key fault parameters are accurately located. The system can automatically trigger hierarchical early warnings (such as switching to a standby magnetic core, emergency shutdown), reduce the equipment maintenance cost, and extend the service life of the magnetic core; at the same time, ensure the stability of the plasma concentration and improve the consistency of the etching rate in the semiconductor process. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0016] Figure 1 Shows the structural block diagram of a diagnostic device for the magnetic core demagnetization fault of a plasma source provided by the embodiments of the present invention; Figure 2 Shows the flowchart of a diagnostic method for the magnetic core demagnetization fault of a plasma source provided by the embodiments of the present invention; Figure 3 Shows the flowchart of the training process of the BP-MTN model provided by the embodiments of the present invention; Figure 4 Shows the network structure diagram of the fully connected layer and the activation layer of the multi-dimensional Taylor network in the embodiments of the present invention; Figure 5 Shows the statistical chart of the identification of fault feature nodes in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] The following will describe in detail the embodiments of the technical solutions of the present invention with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and thus are only examples and cannot be used to limit the protection scope of the present invention. It should be noted that unless otherwise specified, the technical terms or scientific terms used in this application should have the ordinary meaning understood by those skilled in the art to which the present invention belongs.

[0018] Figure 1 The structural block diagram of a diagnostic device for the demagnetization fault of a plasma source magnetic core provided by an embodiment of the present invention. This device is used to implement a diagnostic method for the demagnetization fault of a plasma source magnetic core based on a multi-dimensional Taylor network and a BP neural network. Refer to Figure 1 , the plasma source system includes a reaction chamber, a full-bridge inverter circuit, and an LCL resonant converter. The full-bridge inverter circuit is composed of switching tubes S1 - S4; Vin is an input DC voltage source; the inductor Lr, Lk, and capacitor Cr form an LCL resonant network; the reaction chamber includes an air inlet / outlet and a ferrite magnetic core assembly; this diagnostic device includes: a sensor sampling unit, a dissociation rate detection unit, and a multi-dimensional Taylor network neural network module; the sensor sampling unit includes a voltage sensor for obtaining the ignition voltage Vloop and a temperature sensor for obtaining the temperature data inside the reaction chamber; The dissociation rate detection unit includes an infrared spectrometer, which is used to detect the outlet gas spectrum and compare it with an offline sample database to obtain the gas dissociation rate λ; the multi-dimensional Taylor network neural network module runs a multi-dimensional Taylor network and a BP neural network model (BP-MTN) based on a processor, and takes the average value of the temperature change rate of the magnetic core per unit time, the ignition voltage Vloop, and the gas dissociation rate λ as inputs, and the output is the demagnetization degree d of the magnetic core; a fault warning is triggered according to the value of the output demagnetization degree d.

[0019] Among them, the BP neural network is a multi-layer feedforward network. It consists of three layers: an input layer, a hidden layer, and an output layer. During the implementation process, data is transmitted from the input layer to the output layer. The hidden layer mainly non-linearly transforms the data transmitted by the input layer into a feature matrix, and then the output layer calculates the error with the expected data, and propagates the error forward from the output layer. The connection weights between each layer of the network and the neuron biases of each layer are adjusted by the backpropagated error to optimize the entire neural network and achieve the purpose of learning data features.

[0020] The working process of the BP neural network is divided into two parts: the forward transmission of the system and the backpropagation of the error. The forward transmission process is mainly divided into receiving information, processing information, and outputting information; when the output has a large gap with the expected value, the system will start to backpropagate the error, and each weight value will be changed and adjusted according to the actual situation of the error until the output meets the requirements or reaches the set number of iterations, and the training of the system will end; Utilize the self-learning ability of the BP neural network to online adjust the weight value of the value function, preprocess to obtain the nearest weight control parameter, and input it to the multi-dimensional Taylor network module.

[0021] Figure 2 The step flowchart of a diagnostic method for the demagnetization fault of a plasma source magnetic core provided by an embodiment of the present invention. Refer to Figure 2 , this method includes the following steps: A: Data acquisition: Real-time obtain the temperature change rate ΔT per unit time of the magnetic core, the ignition voltage Vloop, and the gas dissociation rate λ; The temperature of the ferrite magnetic core at time k Is detected by a temperature sensor, and the temperature of the magnetic core detected after one sampling period is ; Then the temperature change rate per unit sampling period is: ; The ignition voltage Vloop is detected by a voltage sensor; For the gas dissociation rate λ, it is necessary to pay attention to the plasma load monitoring signal. The change in the cavity plasma load includes the gas flow rate and pressure injected into the cavity, as well as the gas flow rate at the plasma outlet; then it is detected by an infrared spectrometer at the gas outlet, and the gas dissociation rate λ is obtained by comparing the spectrum with the sample database, where the spectrum corresponding to different dissociation rates is in the sample database.

[0022] B: Data preprocessing: Normalize the temperature change rate ΔT, the ignition voltage Vloop, and the gas dissociation rate λ; Normalize the temperature change rate ΔT, the ignition voltage Vloop, and the gas dissociation rate λ according to historical data; C: Input the data into the trained BP-MTN model and output the demagnetization degree d; The input temperature change rate ΔT, ignition voltage Vloop, and gas dissociation rate λ data are the mean values per unit time.

[0023] D: Estimate the demagnetization degree of the magnetic core according to the demagnetization degree d and process them separately.

[0024] The fault classification is shown in Table 1 below. The demagnetization degree is divided into five levels: 0 (no demagnetization), 0.2 (slight), 0.5 (moderate), 0.8 (severe), and 1 (complete demagnetization) according to the demagnetization degree d, and they are processed separately.

[0025]

[0026] As Figure 3 shown, in this embodiment, the training process of the BP-MTN model is as follows: Step 1: Construct a sample data set according to historical test data: The temperature change rate per unit time of the magnetic core when the remote plasma source is ignited and the ignition voltage Vloop and the gas dissociation rate λ; then, use the temperature change rate per unit time of the magnetic core based on , the ignition voltage V, and the gas dissociation rate λ to construct a sample data set {(x, y, z)}, where x represents the one-dimensional set vector of the temperature change rate per unit time of the magnetic core and y represents the one-dimensional set vector of the ignition voltage Vloop , , z represents a one-dimensional set vector of the gas dissociation rate λ ; Step 2: Construct the x, y, and z in the sample data set {(x, y, z)} into a two-dimensional input matrix, and then perform data normalization operation and update on the two-dimensional input matrix; at the same time, establish the output sample data set d{(x, y, z)} in which d is the label vector corresponding to the network predicted value of the demagnetization degree of the ferrite core; The temperature change rate of the core per unit time , the average value of the ignition voltage Vloop, and the average value of the gas dissociation rate λ are used as the inputs of the BP neural network, and the corresponding demagnetization degree is used as the output set of the BP neural network, and a BP neural network diagnosis model d = F( , V, λ) is constructed, where d is the different demagnetization degrees, is the average value of the temperature change rate of the core per unit time of, V is the average value of the ignition voltage Vloop, and λ is the average value of the gas dissociation rate. According to the demagnetization degree d, real-time online diagnosis of the demagnetization fault of the ferrite core is realized.

[0027] Normalize all eigenvalue features of the sample data set {(x, y, z)}: The normalization process is to uniformly transform each resonant current eigenvalue sequence to within the interval, and the calculation formula is: (1); Among them, Mi is the input eigenvalue after normalization, Ni is the input eigenvalue before normalization, Nmax is the maximum input eigenvalue, and Nmin is the minimum input eigenvalue.

[0028] Step 3: Use the two-dimensional input matrix updated by the normalization operation to train the BP-MTN model and output the network predicted value of the demagnetization degree d of the ferrite core At the same time, initialize the weight set vector on the two-dimensional output node of the BP-MTN model , the bias b of the final output node of the BP-MTN model, and then calculate the error between the label vector d and the network predicted output value ; The structure of the BP network adopts a structure including an input layer, a hidden layer, and an output layer; where i represents the input layer node, j represents the hidden layer node, l represents the output layer node, and determine the number of input nodes m, the number of hidden layer nodes q, and the number of output layer nodes is 3; and initialize the initial values of the weighted coefficients from the input layer to the hidden layer and from the hidden layer to the output layer , and and the offset; the calculation formula for each layer is: (2); where T(x) represents the activation function 1, represents the weight coefficient connecting the i-th neuron in the previous layer and the j-th neuron in the current layer, and b represents the offset of the j-th neuron in the current layer; For a multi-layer network, the feed-forward propagation method is used for calculation, that is, each layer is calculated according to formula (2) until the last output layer; for the i-th neuron in the input layer, its output is the i-th eigenvalue of the input data.

[0029] Step S4: In the forward propagation process, after the input data is calculated by the perceptron node, it is processed by the activation function T(x) to obtain the output result; the activation function T(x) of the hidden layer is selected as the ReLU function, and the activation function l of the BP neural network: (3); Then the output of the forward propagation is: (4).

[0030] Step S5: Calculate the overall error E for the entire sample set, and determine whether the error E meets the design requirements. If it meets, the neural network training ends; if it does not meet, use backpropagation for neural network learning, compare the output result with the expected structure, and according to the error convergence curve during training, fit the demagnetization degree D(k) of the ferrite core of the output quantity through the Taylor polynomial expansion of the activation function 2, and iteratively adjust the weighting coefficients and , online multiple times to implement the weights and the offset b, so that the fitting result approaches the D(k) value at time t, thereby realizing the adaptive adjustment of the control parameters; the multi-dimensional Taylor network MTN model at this time is the BP-MTN model. In this embodiment, the network structure diagrams of the fully connected layer and the activation layer of the multi-dimensional Taylor network are as shown in Figure 4 shown.

[0031] The error function of the BP neural network is the mean square error function: (5); Solving formula (5) for a set of W and b to minimize the error function of the BP neural network; where W is the weight, b is the threshold parameter, m is the number of training samples, k is the number of outputs, is the predicted value of the j-th output of the i-th sample, is the corresponding true value.

[0032] Suppose the input layer of the multi-dimensional Taylor network has n nodes, and the data processing layer has N(n, m) nodes. The data processing layer realizes the weighted summation of the product terms of each power of the input variables; The multi-dimensional Taylor network uses a polynomial composed of addition and multiplication to approximate non-linear functions and can be used to fit multi-variable functions. Therefore, the multi-dimensional Taylor network can be applied to the fault diagnosis of ferrite cores. The basic principle of the multi-dimensional Taylor network is as follows: Considering that the system dynamics equation generally has the following form: (6); In the formula, x ∈ , d ∈ , y ∈ , and f and g are non-linear mapping systems. Equation (6) can be transformed into the following form: (7).

[0033] An important theoretical basis for automatic control is feedback control, that is, the control signal is determined according to the deviation between the given quantity and the feedback quantity. According to the Weierstrass approximation theorem and the Taylor formula, the optimal control signal can be infinitely approximated by the high-order terms of the deviation signal. Thus, the model of the multi-dimensional Taylor network can be designed, and its mathematical expression form is shown in Equation (8): (8); In the formula, the data processing layer is regarded as N(n, m) nodes, where n is the dimension of the system and the highest number of terms of the highest degree is m; the weight matrix between the data processing layer and the output layer is , represents the weight value before the product term of the t-th variable; The highest number of phases obtained by combination, and e is the error between the demagnetization degree and the expected value: ; e i (t) can be expressed in the following form: (9); The output of the input layer is linearly combined to obtain the output , and the combination coefficient is , and a0 is the initial value: (10); During the backpropagation process, the activation function 2 of the multi-dimensional Taylor network MTN uses the tansig function: (11); The tansig function activation function in Equation (11) is Taylor-expanded, where x represents the specimen data: (12); The Taylor expansion of the activation function T2(x) is described generally, where represents the Peano remainder after expansion: (13); Network input , the network output is , the expected output , after using the Taylor expansion of the activation function of the BP neural network to replace the original activation function, the output of the j-th node in the data processing layer is : (14) The gradient descent method is used to adjust the weight parameters of the neural network. At the same time, an inertia term is added to accelerate the search process and promote rapid convergence to the global minimum, then there is: (15); In the formula, η is the learning rate, and α is the inertia coefficient; Taking the output of the j-th node as an example for backpropagation, the gradient of the network-related weights is calculated according to the activation function 2 method, as shown in the following formula: (16); Among them, t1, t2, and t3 represent the accumulated values obtained by multiplying the products of each item in the input layer by the corresponding weights, and their values are as follows: (17) Among them, m represents the number of times of the highest expansion item in the data processing layer of the multi-dimensional Taylor network, N(2, m) represents the number of terms of the polynomial after the m-th power expansion of the input x(k), y(k), z(k) in the data processing layer of the multi-dimensional Taylor network, w 1j , w 2j , w 3j ; represent the weights corresponding to the three nodes t1, t2, and t3 in the output layer respectively; p j,x , p j,y , p j,z Then represent the powers of x(k), y(k), z(k) in the j-th polynomial; Here, the variable is unknown, but the relative changes of E(k) and d(k) can be measured, that is: (18); This brings computational inaccuracies, which can be compensated by adjusting the learning rate η. Doing so can simplify the operation on the one hand. On the other hand, it avoids the situation where formula (18) tends to infinity when d(k) and d(k - 1) are very close. This substitution is acceptable in the algorithm because is a product factor in formula (18), and the positive or negative sign of it determines the direction of weight value change. While the magnitude of the numerical change only affects the speed of weight value change. However, the speed of weight value change can be adjusted by the learning step size.

[0034] Calculate the average gradient of each weight from the gradient vector obtained from formula (18) 、 、 、 ,and update the weights according to formula (19) as follows; (19).

[0035] Then enter step S6; According to the normal threshold interval of each fault feature data, judge whether there is an abnormality in the predicted state of each fault feature corresponding to the next time of the current time. If so, determine each fault feature with an abnormal predicted state as each abnormal fault feature, and combine the corresponding relationship between each specified type of fault and the involved fault features preset in the ferrite core, and further determine that in the predicted state of the ferrite core corresponding to the next time of the current time, there are each specified type of fault corresponding to each abnormal fault feature; otherwise, determine that the predicted state of the ferrite core corresponding to the next time of the current time is normal. As Figure 5 shown, establish n identification statistical nodes corresponding to the input fault features to facilitate distinguishing the corresponding relationship between the fault features and the fault types, and perform the above-mentioned fault type prediction.

[0036] Step S6: Divide the demagnetization degree d of the ferrite core in the output set into n equal parts, and the step size of each part is Then the demagnetization degree is divided into d1, d2,..., dn from small to large (where d1 = 0, indicating that the ferrite core has no demagnetization fault); let ; For example, first set the first demagnetization degree d1 = 0, that is, the ferrite core has no demagnetization fault; the second demagnetization degree dn / 5 = 0.2, that is, the ferrite core has a slight demagnetization fault; the third demagnetization degree dn / 2 = 0.5, that is, the ferrite core has a medium demagnetization fault; the fourth demagnetization degree d4n / 5 = 0.8, that is, the ferrite core has a serious demagnetization fault; the fifth demagnetization degree dn = 1, that is, the ferrite core has a complete demagnetization fault.

[0037] Step S7: Use the fault identification module to identify the non-demagnetization fault of the ferrite core, count the output results of the neural network, and compare them with the real fault state to count the accuracy rate.

[0038] The embodiments of the present invention have the following technical effects: First, by integrating the BP neural network and the Multidimensional Taylor Network (MTN) model, high-precision dynamic diagnosis of core demagnetization faults is achieved. Compared with traditional single neural network methods, the diagnostic accuracy is improved, and real-time online monitoring is supported.

[0039] Second, based on the normalization process and the adaptive weight adjustment mechanism of the Multidimensional Taylor Network, the convergence speed and stability of the model are significantly improved, the training cycle is shortened, and robustness is maintained under extreme working conditions such as high frequency and high temperature.

[0040] Third, through five-level demagnetization degree classification and fault feature contribution analysis, key fault parameters are accurately located. The system can automatically trigger hierarchical warnings (such as switching to a standby core, emergency shutdown), reducing equipment maintenance costs and extending the service life of the core; at the same time, ensuring the stability of plasma concentration and improving the consistency of the etching rate in the semiconductor process.

[0041] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A diagnostic method for the demagnetization fault of a plasma source magnetic core, characterized in that, It includes the following steps: A: Obtain the temperature change rate ΔT, ignition voltage Vloop, and gas dissociation rate λ of the magnetic core in real time per unit time; B: Perform normalization processing on the temperature change rate ΔT, ignition voltage Vloop, and gas dissociation rate λ; C: Input the data after normalization processing into the trained BP-MTN model to output the demagnetization degree d; D: Estimate the demagnetization degree according to the demagnetization degree d and perform separate processing.

2. The diagnostic method according to claim 1, characterized in that: In step A: The gas dissociation rate λ is obtained by detecting the outlet spectrum with an infrared spectrometer and comparing it with a preset sample database.

3. The diagnostic method according to claim 1, characterized in that: The normalization formula in step B is: ； where Mi is the input eigenvalue after normalization, Ni is the input eigenvalue before normalization, Nmax is the maximum input eigenvalue, and Nmin is the minimum input eigenvalue.

4. The diagnostic method according to claim 1, wherein: The construction and training method of the BP-MTN model in step C is as follows: (i) Construct a BP neural network model; (ii) Calculate the output value d through forward propagation and calculate the error based on the mean square error function; (iii) Perform Taylor expansion on the activation function using a multi-dimensional Taylor network, and dynamically adjust the weights and biases through backpropagation until the error converges.

5. The diagnostic method according to claim 4, wherein: In (iii) of step C, the output layer weight update formula of the multi-dimensional Taylor network is: ； where η is the learning rate, α is the inertia coefficient, and E is the mean square error function.

6. The diagnostic method according to claim 4, characterized in that: In (iii) of step C, analyze the contribution degree of each input parameter to the demagnetization degree through statistical nodes, and preferentially optimize the weights of the key parameters whose contribution degree exceeds the preset threshold.

7. The diagnostic method according to claim 1, characterized in that: Divide the demagnetization degree into five levels according to the output value d.

8. A diagnostic device for the demagnetization fault of a plasma source magnetic core, used to implement the diagnostic method as described in any one of claims 1-7, includes: A sensor sampling unit, including a voltage sensor for obtaining the ignition voltage Vloop and a temperature sensor for obtaining the temperature data in the reaction chamber; A dissociation rate detection unit, including an infrared spectrometer, used to detect the outlet spectrum, compare it with an offline sample database, and obtain the gas dissociation rate λ; Multidimensional Taylor network neural network module, taking the temperature change rate of the magnetic core per unit time , the average value of the ignition voltage Vloop and the gas dissociation rate λ as inputs, and the output is the demagnetization degree d of the magnetic core; estimating the demagnetization degree according to the output demagnetization degree d and processing them separately.

Images