Polishing solution service life online detection method and system based on multi-source data and chemical mechanical polishing

Abstract

The invention belongs to the field of machining, and discloses a polishing solution service life online detection method and system based on multi-source data and chemical mechanical polishing, and the method comprises the steps: collecting the multi-source data information of a polishing solution and a polishing platform in the machining process; performing offline polishing test to obtain detection data and a removal rate change curve during circular polishing of the polishing solution; the edge computing platform receives the sensor data, preprocesses the sensor data and uploads a processing result to the upper computer; inputting the preprocessed multi-source data into a polishing solution service life prediction model, and predicting the residual service life of the polishing solution and the current theoretical material removal rate; a prediction result is fed back to an operator in real time through a control system interface, and the current remaining life of the polishing solution and the predicted time node for replacing the polishing solution are prompted. The state of the polishing solution can be monitored in real time, an off-line detection result does not need to be waited, the continuity and stability of the polishing process are guaranteed, and the process problems caused by aging of the polishing solution are reduced.

Description

Technical Field

[0001] The invention relates to the field of mechanical processing, and in particular to an online detection method and system for polishing liquid life based on multi-source data and chemical mechanical polishing. Background Art

[0002] In the manufacturing process of high-hardness and brittle materials, chemical mechanical polishing (CMP) is a key planarization process in integrated circuit manufacturing, and polishing fluid is one of its core consumables. The performance of the polishing fluid is generally evaluated by measuring the material removal rate (MRR) and surface roughness (Ra). During the polishing process of high-hardness and brittle materials, the chemical composition (such as pH value, oxidant concentration, etc.) and physical properties (such as particle size distribution) of the polishing fluid will change with the increase of time, affecting the polishing removal rate and process stability. When the removal rate is lower than the expected value or the surface roughness increases due to the polishing fluid during the polishing process, it can be considered that the polishing fluid has run out of life and needs to be replaced. The existing technology for judging the life of the polishing fluid mostly relies on the method of manual offline detection of material removal rate or empirical rules, which has problems such as detection lag and low accuracy, which can easily lead to waste of polishing fluid or affect the processing quality. Therefore, accurate prediction of the life of the polishing fluid is crucial to maintaining process stability, reducing material waste, and reducing production costs. Summary of the invention

[0003] The object of the present invention is to provide a method and system for online detection of polishing liquid life based on multi-source data and chemical mechanical polishing, so as to solve the problems raised in the above-mentioned background technology.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] An online detection method for polishing liquid life based on multi-source data and chemical mechanical polishing comprises the following steps:

[0006] Step 1, collecting multi-source data information of the polishing liquid and the polishing platform during the processing by arranging multi-source sensors;

[0007] Step 2, performing an offline polishing test to obtain detection data and a removal rate change curve during polishing with a polishing solution in circulation;

[0008] Step 3: The edge computing platform receives sensor data, performs preprocessing, and uploads the processing results to the host computer;

[0009] Step 4, inputting the preprocessed multi-source data into the polishing liquid life prediction model to predict the remaining life of the polishing liquid and the current theoretical material removal rate;

[0010] Step 5, the prediction results are fed back to the operator in real time through the control system interface, indicating the remaining life of the current polishing liquid and the estimated time node for replacing the polishing liquid.

[0011] Furthermore, in step 1, the multi-source sensor includes multiple ones of a particle size distribution detection sensor, a pH value sensor, an oxidant concentration detection sensor, a polishing head rotation speed sensor, a polishing pad rotation speed sensor, and a polishing head pressure sensor. The particle size distribution detection sensor detects the particle size distribution data of the polishing liquid, the pH value sensor detects the pH value of the polishing liquid, the oxidant concentration detection sensor detects the oxidant concentration of the polishing liquid, the polishing head rotation speed sensor detects the polishing head rotation speed, the polishing head pressure sensor detects the pressure value of the polishing head, and the polishing pad rotation speed sensor detects the polishing pad rotation speed.

[0012] Furthermore, the preprocessing in step 3 includes: the edge computing platform timestamps the collected multi-source data to unify the data time label, and uses a recursive average filtering method to remove noise from the polishing liquid particle size distribution data, polishing liquid pH value data, polishing liquid oxidant concentration data, polishing head speed data, and polishing pad speed data to smooth the data to improve data accuracy and ensure data quality, and standardize all sensor data to the range of 0-1 for model input processing.

[0013] Furthermore, step 4 specifically includes: sampling multi-source sensor data within a period of time, and obtaining material removal rate data within the corresponding period through offline measurement, and building a polishing liquid life prediction model based on the detection data and the polishing liquid under certain speed and pressure conditions through machine learning methods, and determining the remaining life of the polishing liquid according to the corresponding removal rate change curve and the minimum removal rate expected value, wherein the polishing liquid life prediction model selects the "long short-term memory network LSTM-multilayer perceptron MLP hybrid model", and uses the preprocessed multi-source data sequence and the polishing liquid life prediction as input and output, and finally combines the credibility between the predicted value and the actual offline removal rate measurement value to obtain the true polishing liquid life value.

[0014] Furthermore, the network structure of the polishing liquid life prediction model includes four layers: an input layer, an LSTM network layer, an MLP network layer, and an output layer.

[0015] The input layer is responsible for receiving multi-source data pre-processed from the edge computing platform and normalizing the multi-source data into an input format that can be used by the model;

[0016] The LSTM network layer processes the smooth time series of the input layer input, captures the long-term and short-term dependencies of multi-source data, and learns the nonlinear correlation between multi-source data. Its structure includes forget gate, input gate, output gate and cell state:

[0017] Forget gate: decides whether to retain the previous cell state;

[0018] ft =σ(W f ·[s t-1 ,x t ]+b f )

[0019] where f t is the forget gate vector, s t-1 is the state of the hidden unit at the previous moment, x t is the current input; W f is the weight coefficient of the forget gate, b f is the bias of the forget gate, σ is the forget gate activation function sigmoid, f t The range of is in [0,1], if f t The closer the value is to 1, the higher the unit state s is at the previous moment. t-1 The corresponding value should be remembered, if f t The closer the value is to 0, the higher the unit state s is at the previous moment. t-1 The more the corresponding value should be forgotten;

[0020] Input gate: determines whether to update the cell state;

[0021] i t =σ(W i ·[s t-1 ,x t ]+b i )

[0022]

[0023] Among them, i t is the input gate vector at time t, σ is the input gate activation function sigmoid, W i is the bias term of the input gate, s t-1 is the state of the hidden unit at the previous moment, x t is the input at the current moment, b i is the bias of the input gate; is the candidate cell state containing possible new information, tanh is the hyperbolic tangent activation function, W c is the weight coefficient for updating the cell state, b c is the bias of the candidate unit state; c t-1 is the cell state at the previous moment, c t is the cell state at the current moment.

[0024] Output gate: determines whether to output the result of the current cell state.

[0025] o t =σ(W0·[s t-1 ,x t ]+b0)

[0026] s t =o t *tanh(c t )

[0027] Among them, t is the output gate vector at time t, s t-1 is the state of the hidden unit at the previous moment, x t is the input at the current moment, b0 is the bias of the output gate, σ is the output gate activation function sigmoid, s t It is the state of the hidden unit at the current moment and is also the final output.

[0028] The MLP network layer is responsible for inputting the trend features output by the LSTM network layer into the multi-layer perceptron to evaluate the remaining life of the polishing liquid, and the output result is the normalized prediction result.

[0029] The output layer is responsible for inversely transforming the normalized prediction results to restore the actual prediction time value.

[0030] Furthermore, the LSTM network layer chooses to construct a 3-layer LSTM network, each layer contains 128 neurons; the MLP network layer chooses to construct a 4-layer multilayer perceptron, and uses a ReLU activation function to enhance nonlinear expression capabilities.

[0031] Furthermore, the polishing liquid life prediction model is deployed in the host computer, and its input layer divides the preprocessed data into fixed time windows to form time series data blocks to input into the LSTM network layer. The input dimension is the multi-source sensor data in each window. The window collects 15 seconds of data once per second, and the input dimension is 15×6.

[0032] Furthermore, if the network prediction of the polishing liquid life prediction model shows that the polishing liquid theoretical removal rate will be lower than the expected minimum value after time T, the host machine will display that the polishing liquid life will expire after time T and prompt that the polishing liquid needs to be replaced when it expires.

[0033] The present invention also provides an online detection system for polishing liquid life based on multi-source data and chemical mechanical polishing, which is used to implement the detection method as described above, comprising:

[0034] A polishing liquid supply end detection subsystem, including a particle size distribution detection sensor, a pH value sensor, and an oxidant concentration detection sensor, wherein the particle size distribution detection sensor is used to detect the particle size distribution of the polishing liquid, the pH value sensor is used to detect the pH value of the polishing liquid, and the oxidant concentration detection sensor is used to detect the oxidant concentration of the polishing liquid;

[0035] The polishing machine end detection subsystem includes a polishing head speed sensor, a polishing head pressure sensor, and a polishing pad speed sensor. The polishing head speed sensor is used to detect the polishing head speed, the polishing head pressure sensor is used to detect the polishing head pressure value, and the polishing pad speed sensor is used to detect the polishing pad speed;

[0036] An edge computing platform that processes raw data collected by multi-source sensors; and

[0037] The host computer contains the polishing liquid life prediction model, and the host computer is used to process the pre-processed multi-source data, generate prediction results and feed them back to the operator through the system control interface.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows: the traditional method often uses two methods to determine the life of the polishing liquid. One is to measure the material removal rate offline over a period of time and determine the current life of the polishing liquid by the change in the material removal rate; the other is to predict the life of the polishing liquid by the empirical rule method and replace the polishing liquid when it expires. Compared with the traditional method, the detection method provided by the present invention can monitor the state of the polishing liquid in real time without waiting for the offline detection results, thereby ensuring the continuity and stability of the polishing process, reducing process problems caused by aging of the polishing liquid, and improving production efficiency. At the same time, by predicting the life of the polishing liquid, the present invention can avoid premature replacement or excessive use of the polishing liquid, thereby reducing material waste and reducing production costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The present invention is a flow chart of an online detection method for polishing liquid life based on multi-source data and chemical mechanical polishing.

[0040] Figure 2 This is a curve of the removal rate change when the polishing liquid is circulated to polish silicon carbide.

[0041] Figure 3 It is a schematic diagram of the structure of the polishing platform and the polishing liquid supply system in the present invention.

[0042] Figure 4 The present invention is a schematic diagram of the structure of an online detection system for polishing liquid life based on multi-source data and chemical mechanical polishing.

[0043] Marking description: 1. Polishing liquid recovery pipeline, 2. Primary recovery barrel, 3. Liquid supply barrel, 4. Waste liquid barrel, 5. First centrifugal pump, 6. Second centrifugal pump, 7. Third centrifugal pump, 8. Fourth centrifugal pump, 9. Recovery filter, 10. Liquid supply filter, 11. Particle size distribution detection sensor, 12. pH value sensor, 13. Oxidant concentration detection sensor, 14. Polishing head pressure sensor, 15. Polishing head speed sensor, 16. Polishing pad speed sensor 17. First electric valve, 18. Polishing liquid supply pipeline. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] like Figure 1-Figure 3 As shown, the present invention provides an online detection method for polishing liquid life based on multi-source data and chemical mechanical polishing, comprising the following steps:

[0046] Step 1: collect multi-source data information of polishing liquid and polishing platform during processing by arranging multi-source sensors. Specifically including:

[0047] First, a polishing liquid supply end detection subsystem is arranged in the liquid supply pipeline. The polishing liquid supply end detection subsystem includes a particle size distribution detection sensor 11, a pH value sensor 12, and an oxidant concentration detection sensor 13. The particle size distribution detection sensor 11 is used to detect the particle size distribution of the polishing liquid. During the detection, the second centrifugal pump 6 supplies liquid to the particle size distribution detection sensor, and the fourth centrifugal pump 8 transports the detected waste liquid to the waste liquid barrel 4. The pH value sensor 12 is used to detect the pH value of the polishing liquid, and the oxidant concentration detection sensor 13 is used to detect the oxidant concentration of the polishing liquid. Then, a polishing machine end detection subsystem is arranged. The polishing machine end detection subsystem includes a polishing head pressure sensor 14, a polishing head speed sensor 15, and a polishing pad speed sensor 16. The polishing head speed sensor 15 is used to detect the polishing head speed. The polishing head pressure sensor 14 is used to detect the polishing head pressure value, and the polishing pad speed sensor 16 is used to detect the polishing pad speed.

[0048] Step 2: Perform an offline polishing test to obtain the test data and removal rate change curve of the polishing liquid during the polishing cycle. Figure 2 shown.

[0049] Specifically, Figure 3 As shown, taking the polishing process of 6H-SiC wafer to Si surface as an example, the following Figure 3 The polishing platform and polishing liquid supply system shown include a polishing liquid recovery pipeline 1, a primary recovery barrel 2, a liquid supply barrel 3, a waste liquid barrel 4, a first centrifugal pump 5, a second centrifugal pump 6, a third centrifugal pump 7, a fourth centrifugal pump 8, a recovery filter 9, a liquid supply filter 10, a particle size distribution detection sensor 11, a pH value sensor 12, an oxidant concentration detection sensor 13, a polishing head pressure sensor 14, a polishing head speed sensor 15, a polishing pad speed sensor 16, a first electric valve 17, a polishing liquid supply pipeline 18, and a polishing machine. The polishing machine is connected to the primary recovery barrel 2 through the polishing liquid recovery pipeline 1, the primary recovery barrel 2, the first centrifugal pump 5, the recovery filter 9, and the liquid supply barrel 3 are connected in sequence through pipelines, the liquid supply barrel 3 is connected to the second centrifugal pump 6 and the third centrifugal pump 7 through pipelines, the second centrifugal pump 6, the particle size distribution detection sensor 11, the fourth centrifugal pump 8, and the waste liquid barrel 4 are connected in sequence through pipelines, the third centrifugal pump 7, the liquid supply filter 10, the pH sensor 12, the oxidant concentration detection sensor 13, and the first electric valve 17 are connected in sequence through pipelines, and the first electric valve 17 is connected to the polishing machine through the polishing liquid supply pipeline 18.

[0050] The SiO2 substrate polishing liquid for testing is configured, the main components of which are 6%wt of 80nm SiO2 abrasive particles and 4%wt of H2O2, and KOH is used to adjust Ph to 8, and polishing and offline data collection are performed. In this process, the polishing liquid is placed in the liquid supply barrel 3, and is pumped out by the third centrifugal pump 7 through the liquid supply filter 10, the particle size distribution detection sensor 11, the pH value sensor 12, the first electric valve 17 and the polishing liquid supply pipeline 18 to the contact point between the polishing head and the polishing pad in the polishing machine. The excess polishing liquid flows into the polishing liquid recovery pipeline 1 and enters the primary recovery barrel 2. After the polishing liquid in the primary recovery barrel 2 reaches a certain amount, it is pumped out by the first centrifugal pump 5 and re-enters the liquid supply barrel 3 through the recovery filter 9, completing a polishing liquid cycle. When the system set time is reached, the set time in this case is 30 minutes, and the polishing liquid particle size distribution detection needs to be performed, the second centrifugal pump 6 pumps out a trace amount of polishing liquid to the particle size distribution detection sensor 11 for detection, and the waste liquid after detection is pumped out by the fourth centrifugal pump 8 to the waste liquid barrel 4.

[0051] Step 3: The detection subsystem transmits the collected data to the edge computing platform, which pre-processes the data and uploads the processing results to the host computer, including:

[0052] Step 3.1, unify the time tags, take the sensor with the lowest sampling frequency as the time base, and unify the time tags for sensor data with different sampling frequencies.

[0053] Step 3.2, denoising, uses recursive average filtering method to denoise and smooth the input data curve.

[0054] Step 3.3, normalization, normalizes all sensor data to the range of 0-1 for easy model input processing.

[0055] Step 4, sampling multi-source sensor data within a period of time, and obtaining material removal rate data within the corresponding period through offline measurement, and building a polishing liquid life prediction model based on the detection data and the polishing liquid under certain speed and pressure conditions through machine learning methods, and determining the remaining life of the polishing liquid according to the corresponding removal rate change curve and the minimum removal rate expected value, wherein the polishing liquid life prediction model selects the "Long Short-Term Memory Network LSTM-Multilayer Perceptron MLP Hybrid Model", and uses the preprocessed multi-source data sequence and the polishing liquid life prediction as input and output, and finally combines the credibility between the predicted value and the actual offline removal rate measurement value to obtain the true polishing liquid life value. If the network prediction of the polishing liquid life prediction model shows that the theoretical removal rate of the polishing liquid will be lower than the expected minimum value after time T, the upper host will display that the polishing liquid life will expire after time T, and prompt that the polishing liquid needs to be replaced when it expires.

[0056] The network structure of the polishing liquid life prediction model consists of four layers: input layer, LSTM network layer, MLP network layer, and output layer.

[0057] The input layer is responsible for receiving multi-source data pre-processed from the edge computing platform and normalizing the multi-source data into an input format that can be used by the model;

[0058] The LSTM network layer processes the smooth time series of the input layer input, captures the long-term and short-term dependencies of multi-source data, and learns the nonlinear correlation between multi-source data. The LSTM network layer chooses to build a 3-layer LSTM network, each layer contains 128 neurons. Its structure includes forget gate, input gate, output gate and cell state:

[0059] Forget gate: decides whether to retain the previous cell state;

[0060] f t =σ(W f ·[s t-1 ,x t ]+b f )

[0061] where f t is the forget gate vector, s t-1 is the state of the hidden unit at the previous moment, x t is the current input; W f is the weight coefficient of the forget gate, b f is the bias of the forget gate, σ is the forget gate activation function sigmoid, f t The range of is in [0,1], if f tThe closer the value is to 1, the higher the unit state s is at the previous moment. t-1 The corresponding value should be remembered, if f t The closer the value is to 0, the higher the unit state s is at the previous moment. t-1 The more the corresponding value should be forgotten;

[0062] Input gate: determines whether to update the cell state;

[0063] i t =σ(W i [s t-1 ,x t ]+b i )

[0064]

[0065] Among them, i t is the input gate vector at time t, σ is the input gate activation function sigmoid, W i is the bias term of the input gate, s t-1 is the state of the hidden unit at the previous moment, x t is the input at the current moment, b i is the bias of the input gate; is the candidate cell state containing possible new information, tanh is the hyperbolic tangent activation function, W c is the weight coefficient for updating the cell state, b c is the bias of the candidate unit state; c t-1 is the cell state at the previous moment, c t is the cell state at the current moment.

[0066] Output gate: determines whether to output the result of the current cell state.

[0067] o t =σ(W0·[s t-1 ,x t ]+b0)

[0068] s t =o t *tanh(c t )

[0069] Among them, t is the output gate vector at time t, s t-1 is the state of the hidden unit at the previous moment, x t is the input at the current moment, b0 is the bias of the output gate, σ is the output gate activation function sigmoid, s t It is the state of the hidden unit at the current moment and is also the final output.

[0070] The MLP network layer is responsible for inputting the trend features output by the LSTM network layer into the multilayer perceptron to evaluate the remaining life of the polishing liquid, and the output result is the normalized prediction result. The MLP network layer chooses to build a 4-layer multilayer perceptron and uses the ReLU activation function to enhance the nonlinear expression ability.

[0071] The output layer is responsible for inversely transforming the normalized prediction results to restore the actual prediction time value.

[0072] The above-mentioned polishing liquid life prediction model is deployed in the host computer. Its input layer divides the preprocessed data into fixed time windows to form time series data blocks and input them into the LSTM network layer. The input dimension is the multi-source sensor data in each window. The window collects 15 seconds of data once per second, so the input dimension is 15×6.

[0073] Furthermore, the edge computing platform transmits the preprocessed data to the host computer wirelessly and inputs it into the LSTM-MLP hybrid model. The input layer of the LSTM-MLP hybrid model is responsible for dividing the preprocessed data into fixed time windows (such as 15 seconds as a window) to form a time series data block input into the LSTM network layer. The input dimension is the multi-source sensor data in each window. The data is collected in the window for 30 seconds, once per second, and the input dimension is 15×6. The LSTM network layer chooses to build a 3-layer LSTM network, each layer contains 128 neurons, so as to capture the time series characteristics. The output is a trend feature vector containing the multi-dimensional time characteristics of the polishing liquid state. The MLP network layer chooses to build a 4-layer multilayer perceptron (MLP), using the ReLU activation function to enhance the nonlinear expression ability. The input is the output feature of the LSTM, and the output is the predicted value of the remaining life of the polishing liquid. The output layer uses linear activation, and the mean absolute error MAE is calculated between the predicted output and the target output as the loss function. MAE is a clear measure of the average error amplitude and is suitable for the average error comparison of the model. The hidden layer adjusts the total weights in the gradient back propagation formula calculated by the loss function, and uses the Adam algorithm to generate optimization parameters for each iterative learning to make the loss function converge. After the model training is completed, the output layer performs denormalization and other processing on the results and outputs the predicted value at the current moment.

[0074] Step 5, repeating the polishing process, and using the polishing liquid life online detection system to achieve online measurement of the polishing liquid life and the expected removal rate.

[0075] Reference Figure 4As shown, an online detection system for polishing liquid life based on multi-source data and chemical mechanical polishing is used to implement the detection method as described above, including a polishing liquid supply end detection subsystem, a polishing machine end detection subsystem, an edge computing platform, and a host computer. The polishing liquid supply end detection subsystem includes a particle size distribution detection sensor 11, a pH value sensor 12, and an oxidant concentration detection sensor 13. The particle size distribution detection sensor 11 is used to detect the particle size distribution of the polishing liquid, the pH value sensor 12 is used to detect the pH value of the polishing liquid, and the oxidant concentration detection sensor 13 is used to detect the oxidant concentration of the polishing liquid. The polishing machine end detection subsystem includes a polishing head pressure sensor 14, a polishing head speed sensor 15, and a polishing pad speed sensor 16. The polishing head speed sensor 15 is used to detect the polishing head speed, the polishing head pressure sensor 14 is used to detect the polishing head pressure value, and the polishing pad speed sensor 16 is used to detect the polishing pad speed. The edge computing platform processes the raw data collected by the multi-source sensors. The host computer is used to process the pre-processed multi-source data, generate prediction results and feed them back to the operator through the system control interface. The above-mentioned polishing liquid life prediction model is stored in the host computer.

[0076] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An online detection method for polishing liquid life based on multi-source data and chemical mechanical polishing, characterized in that: The steps include: Step 1, collecting multi-source data information of the polishing liquid and the polishing platform during the processing by arranging multi-source sensors; Step 2, performing an offline polishing test to obtain detection data and a removal rate change curve during polishing with a polishing solution in circulation; Step 3: The edge computing platform receives sensor data, performs preprocessing, and uploads the processing results to the host computer; Step 4, inputting the preprocessed multi-source data into the polishing liquid life prediction model to predict the remaining life of the polishing liquid and the current theoretical material removal rate; Step 5, the prediction results are fed back to the operator in real time through the control system interface, indicating the remaining life of the current polishing liquid and the estimated time node for replacing the polishing liquid.

2. The method for online detection of polishing liquid life based on multi-source data and chemical mechanical polishing according to claim 1, characterized in that: In step 1, the multi-source sensor includes a particle size distribution detection sensor, a pH value sensor, an oxidant concentration detection sensor, a polishing head rotation speed sensor, a polishing pad rotation speed sensor, and a polishing head pressure sensor. The particle size distribution detection sensor detects the particle size distribution data of the polishing liquid, the pH value sensor detects the pH value of the polishing liquid, the oxidant concentration detection sensor detects the oxidant concentration of the polishing liquid, the polishing head rotation speed sensor detects the polishing head rotation speed, the polishing head pressure sensor detects the pressure value of the polishing head, and the polishing pad rotation speed sensor detects the polishing pad rotation speed.

3. The method for online detection of polishing liquid life based on multi-source data and chemical mechanical polishing according to claim 2, characterized in that: The preprocessing in step 3 includes: the edge computing platform timestamps the collected multi-source data to unify the data time label, and uses the recursive average filtering method to remove noise from the polishing liquid particle size distribution data, polishing liquid pH value data, polishing liquid oxidant concentration data, polishing head speed data, and polishing pad speed data to smooth the data to improve the accuracy of the data and ensure the quality of the data, and standardize all sensor data to the range of 0-1 for model input processing.

4. The method for online detection of polishing liquid life based on multi-source data and chemical mechanical polishing according to claim 1, characterized in that: The step 4 specifically includes: sampling multi-source sensor data within a period of time, and obtaining material removal rate data within the corresponding period by offline measurement, and building a polishing liquid life prediction model based on the detection data and the polishing liquid under certain speed and pressure conditions by machine learning methods, and determining the remaining life of the polishing liquid according to the corresponding removal rate change curve and the minimum removal rate expected value, wherein the polishing liquid life prediction model selects the "long short-term memory network LSTM-multilayer perceptron MLP hybrid model", and uses the preprocessed multi-source data sequence and the polishing liquid life prediction as input and output, and finally combines the credibility between the predicted value and the actual offline removal rate measurement value to obtain the true polishing liquid life value.

5. The method for online detection of polishing liquid life based on multi-source data and chemical mechanical polishing according to claim 4, characterized in that: The network structure of the polishing liquid life prediction model includes four layers: input layer, LSTM network layer, MLP network layer, and output layer. The input layer is responsible for receiving multi-source data pre-processed from the edge computing platform and normalizing the multi-source data into an input format that can be used by the model; The LSTM network layer processes the smooth time series of the input layer input, captures the long-term and short-term dependencies of multi-source data, and learns the nonlinear correlation between multi-source data. Its structure includes forget gate, input gate, output gate and cell state: Forget gate: decides whether to retain the previous cell state; f t =σ(W f ·[s t-1 ,x t ]+b f ) where f t is the forget gate vector, s t-1 is the state of the hidden unit at the previous moment, x t is the current input; W f is the weight coefficient of the forget gate, b f is the bias of the forget gate, σ is the forget gate activation function sigmoid, f t The range of is in [0,1], if f t The closer the value is to 1, the higher the cell state s is at the previous moment. t-1 The corresponding value should be remembered, if f t The closer the value is to 0, the higher the unit state s is at the previous moment. t-1 The more the corresponding value should be forgotten; Input gate: determines whether to update the cell state; I t =σ(W i ·[s t-1 ,x t ]+b i ) Among them, i t is the input gate vector at time t, σ is the input gate activation function sigmoid, W i is the bias term of the input gate, s t-1 is the state of the hidden unit at the previous moment, x t is the input at the current moment, b i is the bias of the input gate; is the candidate cell state containing possible new information, tanh is the hyperbolic tangent activation function, W c is the weight coefficient for updating the cell state, b c is the bias of the candidate unit state; c t-1 is the cell state at the previous moment, c t is the cell state at the current moment. Output gate: determines whether to output the result of the current cell state. about t =σ(W0 [s t-1 ,x t ]+b0) S t =o t *tanh(c t ) Among them, t is the output gate vector at time t, s t-1 is the state of the hidden unit at the previous moment, x t is the input at the current moment, b0 is the bias of the output gate, σ is the output gate activation function sigmoid, s t It is the state of the hidden unit at the current moment and is also the final output. The MLP network layer is responsible for inputting the trend features output by the LSTM network layer into the multi-layer perceptron to evaluate the remaining life of the polishing liquid, and the output result is the normalized prediction result. The output layer is responsible for inversely transforming the normalized prediction results to restore the actual prediction time value.

6. The method for online detection of polishing liquid life based on multi-source data and chemical mechanical polishing according to claim 5, characterized in that: The LSTM network layer chooses to build a 3-layer LSTM network, each layer contains 128 neurons; the MLP network layer chooses to build a 4-layer multilayer perceptron, and uses the ReLU activation function to enhance the nonlinear expression ability.

7. The method for online detection of polishing liquid life based on multi-source data and chemical mechanical polishing according to claim 5, characterized in that: The polishing liquid life prediction model is deployed in the host computer. Its input layer divides the preprocessed data into fixed time windows to form time series data blocks and input them into the LSTM network layer. The input dimension is the multi-source sensor data in each window. The window collects 15 seconds of data once per second, so the input dimension is 15×6.

8. The method for online detection of polishing liquid life based on multi-source data and chemical mechanical polishing according to claim 4, characterized in that: If the network prediction of the polishing liquid life prediction model shows that the polishing liquid theoretical removal rate will be lower than the expected minimum value after time T, the host machine will display that the polishing liquid life will expire after time T and prompt that the polishing liquid needs to be replaced when it expires.

9. An online detection system for polishing liquid life based on multi-source data and chemical mechanical polishing, used to implement the detection method as described in any one of claims 1 to 8, characterized in that: include: A polishing liquid supply end detection subsystem, including a particle size distribution detection sensor, a pH value sensor, and an oxidant concentration detection sensor, wherein the particle size distribution detection sensor is used to detect the particle size distribution of the polishing liquid, the pH value sensor is used to detect the pH value of the polishing liquid, and the oxidant concentration detection sensor is used to detect the oxidant concentration of the polishing liquid; The polishing machine end detection subsystem includes a polishing head speed sensor, a polishing head pressure sensor, and a polishing pad speed sensor. The polishing head speed sensor is used to detect the polishing head speed, the polishing head pressure sensor is used to detect the polishing head pressure value, and the polishing pad speed sensor is used to detect the polishing pad speed; An edge computing platform, which processes raw data collected by multi-source sensors; as well as The host computer contains the polishing liquid life prediction model, and the host computer is used to process the pre-processed multi-source data, generate prediction results and feed them back to the operator through the system control interface.

Images