Signal processing method and device and model training method and device

Abstract

The invention relates to the technical field of storage, and provides a signal processing method and device and a model training method and device, which are used for reducing the influence of nonlinear noise on a read-back signal and improving the signal quality of the read-back signal. The method specifically comprises the following steps: acquiring a first read-back signal, wherein the first read-back signal is an induction signal corresponding to read operation; inputting a first sequence and the first read-back signal into a first neural network model to obtain a second sequence output by the first neural network model, the first neural network model being used for performing noise reduction processing on the first read-back signal, and the first sequence being a sequence obtained by performing sequence detection on a third sequence, the third sequence is a sequence output by the first neural network model based on the first read-back signal; performing sequence detection on the second sequence to obtain a fourth sequence; and acquiring first data corresponding to the first read-back signal according to the fourth sequence.

Description

Technical Field

[0001] The present application relates to the field of storage technology, and in particular to a signal processing method and device and a model training method and device. Background Art

[0002] In recent years, data storage media represented by high-density optical disks have become an indispensable information carrier in the contemporary information society and are widely used in the fields of information distribution, data archiving and data backup. Data reading of various storage media (such as optical disks, magnetic disks, tapes or solid-state hard disks) is still based on the early technology of the data storage industry. Figure 1 FIG. 2 shows a schematic diagram of the process of reading CD information. Figure 1 As shown, the optical head performs a read operation on the optical disc to obtain a readback signal, and the decoding module processes the readback signal after passing through the signal amplification module and the analog-to-digital conversion module to obtain data corresponding to the readback signal. In order to improve the quality of the readback signal, a linear equalizer such as a finite impulse response filter, a zero-forcing equalizer, a partial response equalizer based on the minimum mean square error criterion, or an adaptive linear equalizer based on a minimum square error algorithm is usually used to reduce the linear noise in the readback signal.

[0003] In the process of acquiring the readback signal, the readback signal of the current record bit (i.e., the position in the optical disc corresponding to the data to be sent) may be affected by multiple record bits adjacent to the current record bit, thereby generating strong inter-symbol interference (or inter-code crosstalk). In addition, there is the influence of medium noise in the process of writing data to the optical disc medium, as well as the influence of various system noises such as circuit systems and mechanical vibrations of the drive, resulting in the readback signal acquired by the optical head being a superposition of multiple nonlinear noises. These nonlinear noises will cause errors in the subsequent signal processing (such as sampling judgment, etc.) of the readback signal, greatly affecting the signal quality of the readback signal. The linear equalizer mentioned above actually reduces the linear noise in the readback signal, but cannot reduce the nonlinear noise in the readback signal. It can be seen that how to reduce the impact of nonlinear noise on the readback signal is a problem that needs to be solved. Summary of the invention

[0004] The embodiments of the present application provide a signal processing method and device and a model training method and device for reducing the influence of nonlinear noise on a readback signal and improving the signal quality of the readback signal.

[0005] In a first aspect, an embodiment of the present application provides a signal processing method, which can be executed by an electronic device, or by a component (such as a chip, etc.) in an electronic device, without limitation. The method specifically includes: obtaining a first readback signal, which is a sensing signal corresponding to a read operation; inputting a first sequence and a first readback signal into a first neural network model to obtain a second sequence output by the first neural network model, the first neural network model is used to perform noise reduction processing on the first readback signal, the first sequence is a sequence obtained by performing sequence detection on a third sequence, and the third sequence is a sequence output by the first neural network model based on the first readback signal; performing sequence detection on the second sequence to obtain a fourth sequence; and obtaining first data corresponding to the first readback signal according to the fourth sequence.

[0006] The neural network model is nonlinear and can learn the nonlinear characteristics in the readback signal. In the embodiment of the present application, the first readback signal is subjected to noise reduction processing by the first neural network model, which can reduce the influence of nonlinear noise on the first readback signal. In addition, linear noise is a special nonlinear noise, so the embodiment of the present application can also reduce the influence of linear noise on the first readback signal. Moreover, the input of the first neural network model includes not only the first readback signal, but also a first sequence, which is a sequence obtained by performing sequence detection on the output of the first neural network model. In this way, the first readback signal is subjected to noise reduction processing step by step and iteratively, which can retain the original characteristics corresponding to the first readback signal, thereby reducing the probability of feature loss of the first readback signal due to noise reduction processing, reducing the detection bit error rate, and improving the signal quality of the first readback signal.

[0007] In a possible implementation, obtaining the first readback signal may specifically include: receiving a read instruction, where the read instruction is used to read first data in a first storage device; and, in response to the read instruction, performing a read operation on the first storage device to obtain the first readback signal.

[0008] In one possible implementation, the first neural network model corresponds to the parameters of the first storage device, such as storing the first neural network model using the model information of the first storage device as an index; or, the sample set used to train the first neural network model corresponds to the parameters of the first storage device, so that the parameters of different storage devices correspond to different neural network models, which can improve reliability.

[0009] In a possible implementation, the first sequence and the first readback signal are input into the first neural network model to obtain the second sequence output by the first neural network model, which can be specifically implemented in the following two ways.

[0010] Mode 1, the number of first sequences is one, the number of third sequences is one, the first readback signal is input into the first neural network model to obtain the third sequence output by the first neural network model; sequence detection is performed on the third sequence to obtain the first sequence; the first sequence and the first readback signal are input into the first neural network model to obtain the second sequence. Optionally, the first readback signal is input into the first neural network model to obtain the third sequence output by the first neural network model, which can be specifically: the first readback signal and the all-zero sequence are input into the first neural network model to obtain the third sequence.

[0011] Mode 2, the number of first sequences is N, the number of third sequences is N, and N is an integer greater than 1; the first readback signal is input into the first neural network model to obtain the first third sequence of the N third sequences output by the first neural network model; sequence detection is performed on the first third sequence to obtain the first first sequence of the N first sequences; for the value of i ranging from 1 to N-1, the following steps are performed: the first readback signal and the i-th first sequence of the N first sequences are input into the first neural network model to obtain the i+1-th third sequence of the third sequence output by the first neural network model; sequence detection is performed on the i+1-th third sequence to obtain the i+1-th first sequence of the N first sequences; wherein i is a positive integer less than N; the first readback signal and the N-th first sequence of the N first sequences are input into the first neural network model to obtain the second sequence. Optionally, the first readback signal is input into the first neural network model to obtain the first third sequence of the N third sequences output by the first neural network model, which can be specifically: the first readback signal and the all-zero sequence are input into the first neural network model to obtain the first third sequence.

[0012] In a possible implementation, the number of the first sequences is determined during the training process of the first neural network model.

[0013] In one possible implementation, when the signal quality calculated based on the fourth sequence and the second sequence is greater than or equal to the first threshold, in other words, when the signal quality calculated based on the sequence before sequence detection and the sequence after sequence detection is greater than or equal to the first threshold, stop inputting the sequence after sequence detection and the first readback signal into the first neural network model.

[0014] In a possible implementation, the method may further include: obtaining M training samples, where the M training samples are sensing signals obtained by performing a read operation on a storage device, and M is a positive integer; for values ​​of j ranging from 1 to M, performing the following steps to obtain a first neural network model: inputting the fifth sequence and the j-th training sample among the M training samples into a second neural network model to obtain a sixth sequence output by the second neural network model, wherein the fifth sequence is a sequence obtained by performing sequence detection on the seventh sequence, and the seventh sequence is a sequence output by the second neural network model based on the j-th training sample; when the sixth sequence meets a training termination condition, a trained second neural network model is obtained; wherein the second neural network model is an initial neural network model or a neural network model obtained by training based on the j-1-th training sample among the M training samples, and j is an integer greater than 0 and less than or equal to M.

[0015] In the above implementation, for each training sample, the training sample and the training detection result corresponding to the output of the neural network model during the training process are used as input for model training, which can retain the original characteristics of the training sample in the process of reducing nonlinear noise and linear noise and reduce the detection bit error rate.

[0016] In a possible implementation, there are multiple storage devices, and parameters of each storage device in the multiple storage devices are the same as parameters of the first storage device, and the first storage device is used to store the first data.

[0017] In one possible implementation, the activation function used in the hidden layer of the second neural network model is a nonlinear activation function, such as a tanh function, which can better learn the nonlinear features in the training samples and help reduce nonlinear noise and linear noise in the readback signal.

[0018] In a second aspect, an embodiment of the present application provides a model training method, which can be executed by an electronic device, or by a component (such as a chip, etc.) in an electronic device. In the method, a second readback signal is obtained, and the second readback signal is a sensing signal obtained by performing a read operation on a storage device; a fifth sequence and a second readback signal are input into a second neural network model to obtain a sixth sequence output by the second neural network model, wherein the fifth sequence is a sequence obtained by performing sequence detection on a seventh sequence, and the seventh sequence is a sequence output by the second neural network model based on the second readback signal; and when the sixth sequence meets the training termination condition, a trained second neural network model is obtained, and the trained second neural network model is used to perform noise reduction processing on the readback signal.

[0019] In a possible implementation, the activation function used by the hidden layer of the second neural network model is a nonlinear activation function.

[0020] In a third aspect, an embodiment of the present application further provides a signal processing device, comprising a noise reduction module and a decoding module; wherein: the noise reduction module is used to obtain a first readback signal, the first readback signal being an induction signal corresponding to a read operation; the first sequence and the first readback signal are input into a first neural network model to obtain a second sequence output by the first neural network model, the first neural network model is used to perform noise reduction processing on the first readback signal, the first sequence is a sequence obtained by performing sequence detection on the third sequence, the third sequence is a sequence output by the first neural network model based on the first readback signal; and, sequence detection is performed on the second sequence to obtain a fourth sequence; the decoding module is used to obtain first data corresponding to the first readback signal according to the fourth sequence.

[0021] In a possible implementation, the noise reduction module includes a first submodule and a second submodule, the output of the first submodule is connected to the input of the second submodule, the output of the second submodule is connected to the input of the first submodule, and the output of the second submodule is also connected to the decoding module, the number of first sequences is one, and the number of third sequences is one; when the first readback signal is input into the first neural network model to obtain the third sequence output by the first neural network model: the first submodule is used to input the first readback signal into the first neural network model to obtain the third sequence output by the first neural network model, and send the third sequence to the second submodule; the second submodule is used to receive the third sequence, perform sequence detection on the third sequence to obtain the first sequence, and send the first sequence to the first submodule; the first submodule is also used to receive the first sequence, input the first sequence and the first readback signal into the first neural network model to obtain the second sequence.

[0022] In a possible implementation, the noise reduction module includes a first submodule and a second submodule, the output of the first submodule is connected to the input of the second submodule, the output of the second submodule is connected to the input of the first submodule, and the output of the second submodule is also connected to the decoding module, the number of first sequences is N, the number of third sequences is N, and N is an integer greater than 1; when the first sequence and the first readback signal are input into the first neural network model to obtain the second sequence output by the first neural network model: the first submodule is used to input the first readback signal into the first neural network model to obtain the first third sequence of the N third sequences output by the first neural network model, and send the first third sequence to the second submodule; the second submodule is used to receive the first third sequence, perform sequence detection on the first third sequence, and obtain the first first sequence of the N first sequences , and sends the first first sequence to the first submodule; for the value of i from 1 to N-1: the first submodule is also used to receive the i-th first sequence in the N first sequences, input the first readback signal and the i-th first sequence into the first neural network model, obtain the i+1-th third sequence in the third sequence output by the first neural network model, and send the i+1-th third sequence to the second submodule; the second submodule is also used to receive the i+1-th third sequence, perform sequence detection on the i+1-th third sequence, obtain the i+1-th first sequence in the N first sequences, and send the i+1-th first sequence to the first submodule; wherein i is a positive integer less than N; the first submodule is also used to receive the N-th first sequence in the N first sequences, input the N-th first sequence and the first readback signal into the first neural network model, and obtain the second sequence.

[0023] In a possible implementation, the device may further include an acquisition module and a training module; the acquisition module is used to obtain M training samples, where the M training samples are sensing signals obtained by performing a read operation on a storage device, and M is a positive integer; the training module is used to perform the following steps for values ​​of j from 1 to M to obtain a first neural network model: input the fifth sequence and the jth training sample among the M training samples into the second neural network model to obtain a sixth sequence output by the second neural network model, wherein the fifth sequence is a sequence obtained by performing sequence detection on the seventh sequence, and the seventh sequence is a sequence output by the second neural network model based on the jth training sample; when the sixth sequence meets the training termination condition, a trained second neural network model is obtained; wherein the second neural network model is an initial neural network model or a neural network model obtained by training based on the j-1th training sample among the M training samples, and j is an integer greater than 0 and less than or equal to M.

[0024] In a fourth aspect, an embodiment of the present application further provides a signal processing device, which includes an equalizer, a detector and a decoder, wherein the output of the equalizer is connected to the input of the detector, the output of the detector is connected to the input of the equalizer, and the output of the detector is also connected to the decoder, wherein: the equalizer is used to obtain a first readback signal, input the first readback signal into a first neural network model, obtain a third sequence output by the first neural network model, and send the third sequence to the detector, wherein the first readback signal is an induction signal corresponding to a read operation, and the first neural network model is used to perform noise reduction processing on the first readback signal; the detector is used to receive the third sequence, perform sequence detection on the third sequence, obtain a first sequence, and send the first sequence to the equalizer; the equalizer is also used to receive the first sequence, input the first sequence and the first readback signal into the first neural network model, obtain a second sequence output by the first neural network model, and send the second sequence to the detector; the detector is also used to receive the second sequence, perform sequence detection on the second sequence, obtain a fourth sequence, and send the fourth sequence to the decoder; the decoder is used to obtain first data corresponding to the first readback signal according to the fourth sequence.

[0025] In a fifth aspect, an embodiment of the present application further provides a signal processing device, which includes an equalizer, a detector and a decoder, wherein the output of the equalizer is connected to the input of the detector, the output of the detector is connected to the input of the equalizer, and the output of the detector is also connected to the decoder, wherein: the equalizer is used to obtain a first readback signal, input the first readback signal into the first neural network model, obtain the first third sequence of N third sequences output by the first neural network model, and send the first third sequence to the detector, wherein the first readback signal is an induction signal corresponding to the read operation, and the first neural network model is used to perform noise reduction processing on the first readback signal, and N is an integer greater than 1; the detector is used to receive the first third sequence, perform sequence detection on the first third sequence, obtain the first first sequence of N first sequences, and send the first first sequence to the equalizer; for the value of i from 1 to N-1: the equalizer is also used to receive the i-th sequence of the N first sequences. a first sequence, inputting the first readback signal and the i-th first sequence into the first neural network model, obtaining the i+1-th third sequence in the third sequence output by the first neural network model, and sending the i+1-th third sequence to the detector; the detector is further used to receive the i+1-th third sequence, perform sequence detection on the i+1-th third sequence, obtain the i+1-th first sequence in N first sequences, and send the i+1-th first sequence to the equalizer; wherein i is a positive integer less than N; the equalizer is further used to receive the N-th first sequence in the N first sequences, inputting the N-th first sequence and the first readback signal into the first neural network model, obtaining the second sequence output by the first neural network model, and sending the second sequence to the detector; the detector is further used to receive the second sequence, perform sequence detection on the second sequence to obtain a fourth sequence, and send the fourth sequence to the decoder; the decoder is used to obtain first data corresponding to the first readback signal according to the fourth sequence.

[0026] In a sixth aspect, an embodiment of the present application further provides a model training device, which includes an acquisition module and a training module, wherein: the acquisition module is used to obtain a second readback signal and send the second readback signal to the training module, the second readback signal is an induction signal obtained by performing a read operation on the storage device; the training module is used to receive the second readback signal and input the fifth sequence and the second readback signal into the second neural network model to obtain a sixth sequence output by the second neural network model, wherein the fifth sequence is a sequence obtained by performing sequence detection on the seventh sequence, and the seventh sequence is a sequence output by the second neural network model based on the second readback signal; and when the sixth sequence meets the training termination condition, a trained second neural network model is obtained, and the trained second neural network model is used to perform noise reduction processing on the readback signal.

[0027] In a seventh aspect, an embodiment of the present application further provides an electronic device, which is used to execute any one of the methods provided in the first aspect or the second aspect above.

[0028] In an eighth aspect, an embodiment of the present application further provides a chip, which is used to read a computer program stored in a memory and execute any one of the methods provided in the first or second aspect above.

[0029] In the ninth aspect, an embodiment of the present application further provides a computer-readable storage medium, wherein instructions are stored in the computer-readable storage medium. When the instructions are executed on a processor, the processor executes any one of the methods provided in the first or second aspect above.

[0030] In the tenth aspect, an embodiment of the present application also provides a computer program product comprising instructions, which, when executed by a processor, enables the processor to execute any one of the methods provided in the first or second aspect above.

[0031] In the eleventh aspect, an embodiment of the present application further provides a communication system, which includes the signal processing device described in any one of the third to fifth aspects above, or the model training device described in the sixth aspect above.

[0032] The technical effects that can be achieved in any of the second to tenth aspects mentioned above can refer to the description of the beneficial effects in the first or second aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A schematic diagram of reading optical disc information;

[0034] Figure 2 A schematic diagram of the structure of a signal processing device provided in an embodiment of the present application;

[0035] Figure 3 A schematic diagram of the structure of another signal processing device provided in an embodiment of the present application;

[0036] Figure 4 A schematic diagram of the structure of another signal processing device provided in an embodiment of the present application;

[0037] Figure 5 A schematic diagram of the structure of another signal processing device provided in an embodiment of the present application;

[0038] Figure 6 A flowchart of a signal processing method provided in an embodiment of the present application;

[0039] Figure 7 A schematic diagram of a process for obtaining a fourth sequence provided in an embodiment of the present application;

[0040] Figure 8 A schematic diagram of the structure of a model training device provided in an embodiment of the present application;

[0041] Fig. 9 A flowchart of a model training method provided in an embodiment of the present application;

[0042] Fig.10 A schematic diagram of a tag signal is provided for an embodiment of the present application;

[0043] Fig.11 A schematic diagram of a process for obtaining a trained second neural network model provided in an embodiment of the present application;

[0044] Fig.12 A schematic diagram of model training provided in an embodiment of the present application;

[0045] Fig.13 A schematic diagram of waveform comparison provided in an embodiment of the present application;

[0046] Fig.14 A schematic diagram of checking the bit error rate provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] The embodiments of the present application provide a signal processing method, device, and model training method and device to reduce the impact of nonlinear noise on the readback signal and reduce the probability of feature loss of the readback signal during the noise reduction process, thereby reducing the detection bit error rate and improving the signal quality of the readback signal. In order to make the purpose, technical solution, and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in detail below in conjunction with the accompanying drawings.

[0048] The embodiments of the present application are applicable to the scenario of reading data from a storage device, which may be an optical disk, a magnetic disk, a solid-state storage drive, a glass storage device, or a holographic storage device, etc. The embodiments of the present application do not limit the storage medium of the storage device. Unless otherwise specified, the following description takes the storage device as an optical disk as an example.

[0049] Next, a signal processing device provided in an embodiment of the present application is introduced.

[0050] Figure 2 FIG. 1 is a schematic diagram showing the structure of a signal processing device provided in an embodiment of the present application. Figure 2 As shown, the signal processing device includes an equalizer 201 and a detector 202. Optionally, the signal processing device may also include a decoder 203, that is, the signal processing device provided in the embodiment of the present application may have the function of the decoder 203, or may not have the function of the decoder 203. Figure 2 In order to facilitate understanding, the following description is made by taking the signal processing device having the function of the decoder 203 as an example. Among them, the output of the equalizer 201 is connected to the input of the detector 202, the output of the detector 202 is connected to the input of the equalizer 201, and the output of the detector 202 is also connected to the input of the decoder 203.

[0051] The equalizer 201 can be used to perform nonlinear equalization and linear equalization on the readback signal to reduce the influence of nonlinear noise and linear noise on the readback signal. The equalizer 201 can realize equalization of the readback signal through a neural network model. Exemplarily, the equalizer 201 can be used to obtain a first readback signal; input the first readback signal into the first neural network model or input the first readback signal and the output of the detector 202 into the first neural network model to obtain the output of the first neural network model; and send the output of the first neural network model to the detector 202. Inputting the first readback signal into the first neural network model can be replaced by inputting the sequence corresponding to the first readback signal into the first neural network model. Correspondingly, inputting the first readback signal and the output of the detector 202 into the first neural network model can be replaced by inputting the sequence corresponding to the first readback signal and the output of the detector 202 into the first neural network model.

[0052] The detector 202 can be used to perform sequence detection on the output of the equalizer 201, and can use hard decision or soft decision without limitation. The detector 202 and the decoder 203 can constitute turbo iterative decoding to achieve recovery of recorded data and reduce error propagation in the iterative process. Exemplarily, the detector 202 can be used to receive the output of the equalizer 201, perform sequence detection on the output of the equalizer 201, and send the sequence detection result to the equalizer 201 or send the sequence detection result to the decoder 203.

[0053] The decoder 203, which may also be referred to as a decoder, may be used to perform error correction decoding on the encoded data to restore the original data. Exemplarily, the decoder 203 may be used to receive a sequence detection result, and obtain the original data (such as first data) corresponding to the first readback signal according to the sequence detection result. It should be noted that the data obtained by the decoder 203 may be the original data, or the error between the data and the original data may be less than or equal to a set threshold, without limitation.

[0054] The first readback signal is a sensing signal corresponding to a read operation. The read operation is used to obtain first data in the first storage device. For example, the optical head performs a read operation on the first storage device to obtain the first readback signal. The first neural network model can be used to perform noise reduction processing on the first readback signal, such as reducing or removing nonlinear noise and linear noise in the first readback signal.

[0055] Optionally, the first neural network model corresponds to the parameters of the first storage device; or, the sample set used to train the first neural network model corresponds to the parameters of the first storage device. Taking the parameter as model information as an example, the first neural network model is stored with the model information of the first storage device as an index, so that the equalizer 201 can obtain the first neural network model according to the model information of the first storage device; or, the model information of the storage device corresponding to the sample set used to train the first neural network model is the same as the model information of the first storage device.

[0056] Exemplarily, the signal processing device may further include two buffers, such as first in first out (FIFO) buffers, denoted as FIFO204 and FIFO205. Figure 3 As shown. The two buffers are used to cache sequences (or signals) and send sequences (or signals) to the equalizer 201 with a fixed window. For example, FIFO204 can be used to cache the first readback signal and send the first readback signal to the equalizer 201 with a fixed window. FIFO205 can be used to cache the output of the detector 202 and send the output of the detector 202 to the equalizer 201 with a fixed window. Optionally, the signal processing device may also include a memory 206, which can be used to store at least one neural network model, including the first neural network model. For example, the memory 206 can be used to store at least one neural network model with the parameters of the storage device (such as model information) as an index, so that the equalizer 201 obtains the first neural network model from the memory 206. It should be understood that the at least one neural network model can also be stored by an external memory without limitation. Optionally, the signal processing device may also include a lookup table (LUT) 207, which can be used to quickly find activation functions of nonlinear calculations, etc. The equalizer 201 can quickly find the activation function used for nonlinear calculation by interacting with the LUT 207 to improve the calculation efficiency.

[0057] In one possible implementation, based on Figure 2 or Figure 3The signal processing device shown in the figure may also include but is not limited to one or more of the following: an optical head 208, a preamplifier 209, a filter 210, an analog to digital (AD) converter 211, an automatic gain control (AGC) 212, a timing recovery 213, an interleaver 214, or a deinterleaver 215. Figure 3 As shown, the optical head 208 is connected to the preamplifier 209, the preamplifier 209 is also connected to the filter 210, the filter 210 is also connected to the AD converter 211, the AD converter 211 is also connected to the AGC 212, the AGC 212 is also connected to the timing recovery 213, and the timing recovery 213 is also connected to the input of the equalizer 201. The input of the interleaver 214 is connected to the output of the decoder 203, and the output of the interleaver 214 is connected to the input of the detector 202. The input of the deinterleaver 215 is connected to the output of the detector 202, and the output of the deinterleaver 215 is connected to the input of the decoder 203. The detector 202, the decoder 203, the interleaver 214 and the deinterleaver 215 can also constitute turbo iterative decoding to realize the recovery of recorded data and reduce the error propagation of the iterative process.

[0058] The optical head 208 can be understood as a sensor, which is used to convert the light intensity sensed from the optical disc into an electrical signal. For example, the optical head 208 can be used to perform a read operation on the first storage device to obtain a first readback signal. The preamplifier 209 can be used to amplify the electrical signal output by the optical head 208 to solve the problem that the amplitude of the output electrical signal is small due to the small light intensity sensed by the optical head 208. For example, the preamplifier 209 can be used to amplify the first readback signal. The filter 210 can be used to eliminate (or reduce) the electronic noise introduced by the storage system circuit, the low-frequency noise outside the data storage passband range, or the high-frequency noise outside the data storage passband range, that is, the filter 210 can perform preliminary noise reduction processing on the first readback signal. The AD converter 211 can be used to sample the waveform of the first readback signal and forward the first readback signal to a digital signal (such as the first readback signal is presented in the form of a sequence) for subsequent signal processing. AGC 212 can be used to adjust the amplitude of the signal so that the amplitude of the adjusted signal matches the amplitude expected by the detector 202, which is beneficial for subsequent sequence detection. Timing recovery 213 can be used for synchronization processing, such as synchronization processing of the sampling of the first readback signal, to avoid sampling at the wrong time, which may cause errors in the subsequent processing process or inaccurate subsequent processing results. Interleaver 214 and deinterleaver 215 can be used to detect the scrambling and descrambling of the sequence position in the decoding process to prevent errors in the iterative process of sequence detection and decoding.

[0059] Optionally, the signal processing device may further include a controller, Figure 2 , Figure 3 and Figure 4 The controller may be used to receive a read instruction, such as a read instruction from a user terminal. The read instruction is used to read the first data in the first storage device. The controller is also used to drive the optical head 208 to perform a read operation on the first storage device in response to the read instruction. After the signal processing device obtains the first data, the controller may also be used to send the first data, such as sending the first data to the user terminal.

[0060] As mentioned above, the detector 202 may send the sequence detection result to the equalizer 201, or may send the sequence detection result to the decoder 203. Exemplarily, the detector 202 may send the sequence detection result to the equalizer 201 or send the sequence detection result to the decoder 203 based on the signal quality calculated by the input and output of the detector 202, and / or based on the number of cycles between the equalizer 201 and the detector 202; or, the detector 202 may send the sequence detection result to the equalizer 201 or send the sequence detection result to the decoder 203 based on the number of cycles between the equalizer 201 and the detector 202; or, the detector 202 may send the sequence detection result to the equalizer 201 or send the sequence detection result to the decoder 203 based on the signal quality calculated by the input and output of the detector 202 and the number of cycles between the equalizer 201 and the detector 202.

[0061] For ease of understanding, the sequence detection result sent by the detector 202 to the equalizer 201 is recorded as the first sequence, the sequence detection result sent by the detector 202 to the decoder 203 is recorded as the fourth sequence, and the output of the first neural network model is recorded as the second sequence or the third sequence, where the second sequence is used to obtain the fourth sequence and the third sequence is used to obtain the first sequence.

[0062] In one example, if the signal quality calculated based on the input of detector 202 (i.e., the output of the first neural network model) and the output of detector 202 (i.e., the sequence detection result) is less than a first threshold, detector 202 sends the sequence detection result to equalizer 201; if the signal quality calculated based on the output of the first neural network model and the sequence detection result is greater than or equal to the first threshold, detector 202 sends the sequence detection result to decoder 203.

[0063] In another example, if the number of cycles between the equalizer 201 and the detector 202 does not reach N times, the detector 202 sends the sequence detection result to the equalizer 201; if the number of cycles between the equalizer 201 and the detector 202 reaches N times, the detector 202 sends the sequence detection result to the decoder 203. The number of cycles between the equalizer 201 and the detector 202 can be understood as the number of times the detector 202 sends the sequence detection result to the equalizer 201, or it can also be understood as the number of the first sequence, or it can also be understood as the number of the third sequence. Among them, N can be pre-defined, or it can also be set by the user, or it can also be determined during the training process of the first neural network model, without limitation. The N is a positive integer.

[0064] The following describes the specific implementation process of the signal processing device acquiring the first data for the two cases where N is 1 or greater than 1.

[0065] Case 1: N is 1, that is, the number of first sequences is one, and the number of third sequences is also one. In this case 1, the implementation process of the signal processing device obtaining the first data may include the following steps. Step A1, the equalizer 201 inputs the first readback signal into the first neural network model to obtain the third sequence output by the first neural network model, and sends the third sequence to the detector 202. For example, the equalizer 201 inputs the first readback signal and the all-zero sequence into the first neural network model to obtain the third sequence. Step A2, the detector 202 receives the third sequence, performs sequence detection on the third sequence to obtain the first sequence, and sends the first sequence to the equalizer 201. Step A3, the equalizer 201 inputs the first readback signal and the first sequence into the first neural network model to obtain the second sequence output by the first neural network model, and sends the second sequence to the detector 202. Step A4, the detector 202 receives the second sequence, performs sequence detection on the second sequence to obtain the fourth sequence, and sends the fourth sequence to the decoder 203. Step A5, the decoder 203 receives the fourth sequence, and obtains the first data according to the fourth sequence.

[0066] Case 2: N is greater than 1, that is, the number of first sequences is N, and the number of third sequences is also N. In this case 2, the implementation process of the signal processing device obtaining the first data may include the following steps. Step B1, the equalizer 201 inputs the first readback signal into the first neural network model to obtain the first third sequence of the N third sequences output by the first neural network model, and sends the first third sequence to the detector 202. For example, the equalizer 201 can input the first readback signal and the all-zero sequence into the first neural network model to obtain the first third sequence. Step B2, the detector 202 receives the first third sequence, performs sequence detection on the first third sequence to obtain the first first sequence of the N first sequences, and sends the first first sequence to the equalizer 201. Step B3, for values ​​of i ranging from 1 to N-1, includes the following contents: the equalizer 201 receives the i-th first sequence among the N first sequences, inputs the first readback signal and the i-th first sequence into the first neural network model to obtain the i-th third sequence among the N third sequences, and sends the i+1-th third sequence to the detector 202; the detector 202 receives the i+1-th third sequence, performs sequence detection on the i+1-th third sequence to obtain the i+1-th first sequence among the N first sequences, and sends the i+1-th first sequence to the equalizer 201. Wherein, i is a positive integer less than N. Step B4, the equalizer 201 receives the N-th first sequence among the N first sequences, inputs the first readback signal and the N-th first sequence into the first neural network model to obtain the second sequence output by the first neural network model, and sends the second sequence to the detector 202. Step B5, the detector 202 receives the second sequence, performs sequence detection on the second sequence to obtain a fourth sequence, and sends the fourth sequence to the decoder 203. Step B6: the decoder 203 receives the fourth sequence, and obtains the first data according to the fourth sequence.

[0067] It should be noted that the signal processing device provided in the embodiment of the present application may include a hardware structure, or a software module, or a hardware structure and a software module. Whether one of the above functions is executed in the form of a hardware structure, a software module, or a hardware structure plus a software module depends on the specific application and design constraints of the technical solution.

[0068] Based on the same technical concept as the above-mentioned device embodiment, the present application embodiment provides another structural schematic diagram of a signal processing device. Figure 5As shown, the signal processing device may include a noise reduction module and a decoding module. The noise reduction module may be used to implement the functions of the equalizer 201 and the detector 202. For example, the noise reduction module may be used to obtain a first readback signal, input the first sequence and the first readback signal into the first neural network model to obtain a second sequence output by the first neural network model, perform sequence detection on the second sequence to obtain a fourth sequence, and send the fourth sequence to the decoding module. For details, please refer to the aforementioned content, which will not be repeated. The decoding module may be used to implement the function of the decoder 203. For example, the decoding module may be used to receive the fourth sequence and obtain the first data according to the fourth sequence. For details, please refer to the aforementioned content, which will not be repeated. Optionally, the noise reduction module may be divided into a first submodule and a second submodule, the output of the first submodule is connected to the input of the second submodule, the output of the second submodule is connected to the input of the first submodule, and the output of the second submodule is also connected to the input of the decoding module. That is, the first submodule may be used to implement the function of the equalizer 201, and the second submodule may be used to implement the function of the detector 202. For details, please refer to the aforementioned description of the equalizer 201 and the detector 202, which will not be repeated. It should be noted that the signal processing device may also include other modules, such as an AD conversion module or an AGC module, etc. For details, please refer to the aforementioned Figure 3 and Figure 4 The relevant content will not be repeated here.

[0069] Based on the same technical concept as the above-mentioned device embodiment, the present application embodiment also provides a signal processing method. The method can be executed by an electronic device, or by a component of an electronic device. The electronic device can be the above-mentioned Figures 2 to 5 The signal processing device of any of the embodiments shown in Figures 2 to 5 The signal processing device in any of the embodiments shown in FIG. Figure 6 As shown, the method may include the following contents.

[0070] S61: The signal processing device obtains a first readback signal.

[0071] S61 can be Figures 2 to 4 The equalizer 201 in any of the embodiments shown in the figure is executed, or by Figure 5 The noise reduction module in the illustrated embodiment is executed, for example, by the first submodule. The first readback signal is a sensing signal corresponding to the read operation. For example, the signal processing device may receive the first readback signal. Alternatively, the signal processing device may receive a read instruction, the read instruction being used to read the first data in the first storage device; in response to the read instruction, a read operation is performed on the first storage device to obtain the first readback signal.

[0072] S62: The signal processing device inputs the first sequence and the first read-back signal into the first neural network model to obtain a second sequence output by the first neural network model.

[0073] S62 can be Figures 2 to 4 The equalizer 201 or the equalizer 201 and the detector 202 in any of the embodiments shown in the figure may be performed by Figure 5 The noise reduction module in the illustrated embodiment is executed, for example, by the first submodule and the second submodule. The first sequence is a sequence obtained by performing sequence detection on the third sequence, and the third sequence is a sequence output by the first neural network module based on the first readback signal. For the description of the first sequence and the third sequence, please refer to the aforementioned content and will not be repeated here.

[0074] The number of the first sequence and the number of the third sequence are both N. If N is 1, it means that the signal processing device needs to perform the step of inputting the first sequence and the first readback signal into the first neural network model once. Furthermore, before inputting the first sequence and the first readback signal into the first neural network model, the signal processing device may input the first readback signal into the first neural network model to obtain the third sequence output by the first neural network model, and perform sequence detection on the third sequence to obtain the first sequence. For example, the signal processing device inputs the first readback signal and the all-zero sequence into the first neural network model to obtain the third sequence, such as Figure 7 As shown in (1) in .

[0075] If N is greater than 1, it means that the signal processing device needs to perform the step of inputting the first sequence and the first readback signal into the first neural network model N times, and finally obtain the second sequence output by the first neural network model. Specifically, the signal processing device can input the first readback signal into the first neural network model to obtain the first third sequence of the N third sequences output by the first neural network model; and perform sequence detection on the first third sequence to obtain the first first sequence of the N first sequences. Further, for the value of i ranging from 1 to N-1, the signal processing device performs the following steps: input the first readback signal and the i-th first sequence of the N first sequences into the first neural network model to obtain the i+1-th third sequence in the third sequence output by the first neural network model; perform sequence detection on the i+1-th third sequence to obtain the i+1-th first sequence in the N first sequences. Wherein, i is a positive integer less than N. When the N-th first sequence of the N first sequences output by the first neural network model is obtained, the signal processing device can input the first readback signal and the N-th first sequence into the first neural network model to obtain the second sequence. For example, the signal processing device may input the first readback signal and the all-zero sequence into the first neural network model to obtain the first third sequence, such as Figure 7 As shown in (2) in .

[0076] The number of the first sequences (ie, N) may be predefined, or may be determined during the training process of the first neural model, without limitation.

[0077] S63: The signal processing device performs sequence detection on the second sequence to obtain a fourth sequence.

[0078] S63 can be Figures 2 to 4 The detector 202 in any of the illustrated embodiments may be performed by Figure 5 The noise reduction module in the illustrated embodiment is executed, such as by the second submodule. In one example, the signal processing device can calculate the signal quality based on the two sequences before and after the sequence detection in real time. When the signal quality is greater than or equal to the first threshold, the signal processing device no longer inputs the sequence detection result into the first neural network model, but obtains the first data based on the sequence detection result, that is, executes S64. In this embodiment, the signal quality calculated based on the fourth sequence and the second sequence is greater than or equal to the first threshold. Accordingly, the signal quality calculated based on the third sequence and the first sequence is less than the first threshold. Among them, the first threshold can be predefined or set by the user, without limitation.

[0079] In another example, the signal processing device may record the number of executions of S62 (please refer to the description of the number of cycles of the equalizer 201 and the detector 202). When the number of executions reaches N, the signal processing device no longer inputs the sequence detection result into the first neural network model, but obtains the first data based on the sequence detection result, that is, executes S64. The description of N can be found in the above content and will not be repeated here.

[0080] S64: The signal processing device obtains first data corresponding to the first readback signal according to the fourth sequence.

[0081] S64 can be Figures 2 to 4 The decoder 203 in any of the illustrated embodiments may be executed by Figure 5 The decoding module in the illustrated embodiment is executed, such as the second submodule. For example, the signal processing device may perform operations such as decoding the fourth sequence to obtain the first data. The embodiment of the present application does not limit the specific implementation process of the signal processing device obtaining the first data according to the fourth sequence.

[0082] It should be noted that S64 is an optional step. For example, when the signal processing device does not have a decoding function, the signal processing device does not execute S64, but sends the fourth sequence to the decoding module so that the decoding module obtains the first data according to the fourth sequence.

[0083] Based on the same technical concept as the above embodiments, the present embodiment also provides a model training device for obtaining a first neural network model. Figure 8As shown, the model training device includes an acquisition module and a training module. The acquisition module can be used to obtain M training samples (also called readback signals), and the M training samples are sensing signals obtained by performing a read operation on one or more storage devices, and M is a positive integer. The training module can be used to perform the following steps for the value of j from 1 to M to obtain a first neural network model: input the fifth sequence and the jth training sample in the M training samples into the second neural network model to obtain the sixth sequence output by the second neural network model, wherein the fifth sequence is a sequence obtained by performing sequence detection on the seventh sequence, and the seventh sequence is a sequence output by the second neural network model based on the jth training sample; when the sixth sequence meets the training termination condition, or the number of training times reaches the set number of times, a trained second neural network model is obtained; wherein the second neural network model is an initial neural network model or a neural network model obtained by training based on the j-1th training sample in the M training samples, and j is an integer greater than 0 and less than or equal to M. wherein the parameters of each storage device in the multiple storage devices are the same. For example, the parameters of each storage device in the one or more storage devices are the same as the parameters of the first storage device.

[0084] The training algorithm used by the model training device may be a back propagation algorithm, but is not limited thereto.

[0085] Optionally, the model training device obtains training samples at the granularity of the storage device, so that the parameters of the neural network model determined by the training samples are all global optimal solutions. In this way, when the readback window of the optical head 208 changes, there is no need to redetermine the parameters of the neural network model, which can effectively avoid the system delay caused by repeated calculations.

[0086] Exemplarily, the training module can be divided into a third submodule and a fourth submodule. The output of the third submodule is connected to the input of the fourth submodule, and the output of the fourth submodule is connected to the input of the third submodule. The third submodule can be used to input the training sample or the training sample and the output of the fourth submodule into the neural network model to be trained. The fourth submodule can be used to perform sequence detection on the output of the third submodule.

[0087] It should be pointed out that Figure 8 The model training device shown is Figures 2 to 5 The signal processing device in any of the embodiments shown in the figure may also be integrated into the same chip or chip system without limitation. Figures 2 to 5 The signal processing device in any of the embodiments shown in the figure may also include the acquisition module and the training module. In addition, inputting the training sample into the second neural network model may be replaced by inputting the sequence corresponding to the training sample into the second neural network model.

[0088] Based on the same technical concept as the above-mentioned device embodiment, the present application embodiment also provides a model training method, which can be executed by an electronic device or by a component of an electronic device. The electronic device can be the above-mentioned Figure 8 The model training device in the embodiment shown may include Figure 8 The model training device in the embodiment shown. Take the model training device and one training sample (recorded as the second readback signal) among the M training samples as an example. Fig. 9 As shown, the method may include the following contents.

[0089] S91: The model training device obtains a second readback signal.

[0090] S91 can be Figure 8 The acquisition module in the embodiment shown is executed. For example, the model training device can perform a read operation on the second storage device to obtain the second read-back signal. The original data corresponding to the second read-back signal is recorded as the second data. The specific implementation process of S91 can refer to the specific implementation process of S61, which will not be repeated.

[0091] S92: The model training device inputs the fifth sequence and the second read-back signal into the second neural network model to obtain a sixth sequence output by the second neural network model.

[0092] Among them, the fifth sequence is the sequence obtained by performing sequence detection on the seventh sequence. The seventh sequence is the sequence output by the second neural network model based on the second readback signal. The sixth sequence can be understood as the sequence output by the second neural network model in any training process except the first training process. The seventh sequence can be understood as the sequence output by the second neural network model in any training process except the last training process. For example, during the first training, the model training device can input the second readback signal and the all-zero signal into the second neural network model to obtain the seventh sequence output by the second neural network model. For the convenience of description, the sequence output by the second neural network model in the last training process is referred to as the sixth sequence, and the rest are referred to as the seventh sequence.

[0093] The second neural network model may be an initial neural network model, or may be a neural network model obtained by performing model training on a previous training sample of the second readback signal.

[0094] Exemplarily, the model training device can input the second readback signal into the second neural network model in a sliding form through a fixed window, which can improve the adaptability of the neural network model. The fixed window length is denoted as H, where H is a positive integer. The second neural network model includes two inputs, one input is denoted as S1, and the other input is denoted as S2. It should be understood that S1 corresponds to the sequence corresponding to the second readback signal. S2 corresponds to an all-zero sequence or a fifth sequence. Accordingly, the length of S1 is H, and each sample element of S1 corresponds to a feature of the second readback signal. A bit of the original data (such as the first data) corresponding to the sample element at the center of the window is the label signal (or marker signal) of the second readback signal; the length of S2 is also H. During initialization, S2 is an all-zero sequence, and then S2 is a fifth sequence. The center bit of the window of S2 is also a label signal, such as Fig.10 shown. Fig.10 In the example, H is 11. The label signal is used to restore the original data. Optionally, H is an odd number, so that the feature of the label signal in the training sample is located at the center of the original data. Since the readback signal is centrally symmetrical, the label signal that coincides with the center of the original data is an ideal equalization target, so that the trained neural network model can effectively suppress the crosstalk of the surrounding signals at the center position.

[0095] S93: The model training device determines whether the training termination condition is met.

[0096] If the model training device determines that the training termination condition is not met, the content of S92 is executed; if the model training device determines that the training termination condition is met, the content of S94 is executed. The training termination condition may include, but is not limited to: the difference between the sixth sequence and the sequence corresponding to the second data is less than or equal to the second threshold, or the number of training times reaches the set number of training times. For example, the model training device may determine whether the sixth sequence meets the training termination condition based on the loss function.

[0097] Assume that the number of training times is recorded as T, and T is a positive integer greater than 1, then the number of times S92 needs to be executed is T-1. For example, T is 2, which means that the number of the seventh sequence is one, and the number of the fifth sequence is one. The specific training process is as follows: the model training device inputs the second readback signal and the all-zero sequence into the second neural network model to obtain the seventh sequence output by the second neural network model; performs sequence detection on the seventh sequence to obtain the fifth sequence; inputs the fifth sequence and the second readback signal into the second neural network model to obtain the trained second neural network model, such as Fig.11As shown in (1). For another example, T is greater than 2, which means that the number of seventh sequences is T-1, and the number of fifth sequences is T-1. The specific training process is as follows: the model training device inputs the second readback signal and the all-zero sequence into the second neural network model to obtain the first seventh sequence in the T-1 seventh sequences output by the second neural network model; sequence detection is performed on the first seventh sequence to obtain the T-1th fifth sequence in the T-1 fifth sequences. For k from 1 to T-2, the model training device inputs the kth fifth sequence in the T-1 fifth sequences and the second readback signal into the second neural network model to obtain the k+1th seventh sequence in the T-1 seventh sequences output by the second neural network model, and sequence detection is performed on the k+1th seventh sequence to obtain the k+1th fifth sequence in the T-1 fifth sequences. Further, the model training device inputs the t-1th fifth sequence in the T-1 fifth sequences and the second readback signal into the second neural network model to obtain a trained second neural network model, as shown in FIG. Fig.11 It should be noted that the model training device can determine whether the training termination condition ( Fig.11 It can also be determined whether the training termination condition is met based on whether T reaches the set number of training times, etc., without limitation.

[0098] The activation function used in the hidden layer of the second neural network model can be a nonlinear activation function to learn the nonlinear features in the training samples. The nonlinear activation function can be, for example, a tanh function, without limitation. The output of the tanh function is bounded and can be used as the input of the next hidden layer, and the calculation result of the tanh function is symmetrically distributed with 0 as the center, corresponding to the intensity distribution of the readback signal sensed by the optical head 208, which is conducive to the subsequent cascade calculation of soft information with the detector 202 or directly used as soft information, and can simplify the structure of the neural network model.

[0099] Exemplarily, the activation function used by the hidden layer of the second neural network model is the tanh function, and the loss function used by the second neural network model is the minimum mean square error. Alternatively, the activation function used by the hidden layer of the second neural network model is the tanh function, the loss function used by the second neural network model is the cross entropy, and the activation function used by the output layer of the second neural network model is the sigmoid function. Alternatively, the activation function used by the hidden layer of the second neural network model is the tanh function, and the activation function used by the output layer of the second neural network model is the softmax function.

[0100] The following takes the second neural network model including an input layer, L hidden layers and an output layer as an example to introduce the output of each layer of the second neural network model.

[0101] In one example, the number of neurons in the input layer is 2H, and the output y of this input layer (0) is x, that is, two inputs S1 and S2. For the values of H, S1, and S2, please refer to the foregoing content and will not be elaborated here. The output y of the l-th hidden layer among the L hidden layers (l) can be f(w (l) y (l-1) +b (l) ), where y (l-1) is the output of the (l - 1)-th hidden layer among the L hidden layers, w (l) is the coefficient matrix of the l-th hidden layer, b (l) is the bias coefficient vector of the l-th hidden layer, f() is a non-linear activation function (such as the tanh function), and l is a positive integer less than or equal to L. The number of neurons in the output layer is 1, and the output y of the output layer (L+1) can be w (L+1) y (L) +b (L+1) , where w (L+1) is the coefficient matrix of the output layer, b (L+1) is the bias coefficient vector of the output layer, and y (L) is the output of the L-th hidden layer among the L hidden layers. Training algorithm The output is divided into soft output and hard output. The soft output is y k , as shown in (1) of Fig.12 . The hard output is 1 when y k ≥ threshold, and is 0 when y k < threshold, where threshold is the threshold value.

[0102] In another example, the number of neurons in the input layer is 2H, and the output y of this input layer (0) is x, that is, two inputs S1 and S2. For the values of H, S1, and S2, please refer to the foregoing content and will not be elaborated here. The output y of the l-th hidden layer among the L hidden layers (l) can be f(w (l) y (l-1) +b (l) ), where y (l-1) is the output of the (l - 1)-th hidden layer among the L hidden layers, w (l) is the coefficient matrix of the l-th hidden layer, b (l)is the bias coefficient vector of the lth hidden layer, f() is a nonlinear activation function (such as tanh function), and l is a positive integer less than or equal to L. The number of neurons in the output layer is 1, and the activation function used in the output layer is the softmax function. The output of the output layer is the probability of being judged as 1. The closer the probability is to 1, the closer the classification result is to the written data 0. Conversely, the closer the classification result is to the written data 1. The output of the output layer can be the log-likelihood ratio of the probabilities of the written data being judged as 0 and 1 (i.e., soft decision), or the result of the threshold decision binarization of the log-likelihood ratio (i.e., hard decision). For example, the output P of the output layer is sigmoid(y (l+1) ),like Fig.12 As shown in (2) in the figure. The loss function used by the second neural network model is the cross entropy function. The loss value J is in, represents the estimated value of the output of the second neural network model, 1(·) is The indicator function when it is true, p ik is the probability value when the i-th element belongs to the k-th category. Minimizing J results in the most accurate and reliable classification. Assume that the output layer outputs a probability ranging from 0 to 1, denoted as P1, indicating the possibility that the sample target value is "1", and the probability P0 of the target value being "0" is 1-P1. The soft output LLR(k) is log(P0 / P1), and the hard output When LLR(k)>0, it is 0, and when LLR(k)<0, it is 1. It should be understood that the output P of the output layer in this example is P0 or P1.

[0103] Another example, the number of neurons in the input layer is 2H, and the output y of the input layer is (0) is x, that is, two inputs S1 and S2. Please refer to the above content for H, S1 and S2, which will not be repeated here. The output y of the lth hidden layer in the L hidden layers (l) can be f(w (l) y (l-1) +b (l) ), where y (l-1) is the output of the l-1th hidden layer among the L hidden layers, w (l) is the coefficient matrix of the lth hidden layer, b (l) is the bias coefficient vector of the lth hidden layer, f() is a nonlinear activation function (such as tanh function), and l is a positive integer less than or equal to L. The output layer has 2 neurons, and the output P(y=c|x) of the output layer is softmax(w (c) y (l) +b c ), where softmax() is the softmax function, w (c) is the weight of the output layer, b (c)is the bias coefficient vector of the output layer, c is the total category, and in the embodiment of the present application, c is 2, such as Fig.12 As shown in (3) in . Fig.12 In (3), two outputs are respectively recorded as P0 and P1. Please refer to the description of P0 and P1 in the above example, and will not be repeated here.

[0104] S94: The model training device obtains the trained second neural network model.

[0105] S91 to S94 can be performed by a training module in the model training device. For example, when the sixth sequence meets the training termination condition, the model training device obtains the trained second neural network model.

[0106] At this point, the neural network model training for the second readback signal is completed. If the second readback signal is the last training sample among the M training samples, the trained second neural network model is the first neural network model; if the second readback signal is not the last sample among the M training samples, the model training is performed based on the trained second neural network model and the next training sample. For details, please refer to Fig. 9 The steps in the illustrated embodiment are performed until the training of the neural network model for the last training sample among the M training samples is completed, thereby obtaining the first neural network model.

[0107] The neural network model is nonlinear and can learn the nonlinear features and linear features in the readback signal. In the embodiment of the present application, the first neural network model is used to perform noise reduction processing on the first readback signal, which can reduce the impact of nonlinear noise on the first readback signal. In addition, linear noise is a special nonlinear noise, so the embodiment of the present application can also reduce the impact of linear noise on the first readback signal. Moreover, the input of the first neural network model includes not only the first readback signal, but also a first sequence, which is a sequence obtained by performing sequence detection on the output of the first neural network model. In this way, the first readback signal is subjected to noise reduction processing step by step and iteratively, which can retain the original features corresponding to the first readback signal, thereby reducing the probability of feature loss of the first readback signal due to noise reduction processing, reducing the detection bit error rate, and improving the signal quality of the first readback signal.

[0108] exist Fig.13 In the embodiment, the first data is stored in a sequence corresponding to the first storage device, which is recorded as a nonreturn to zero inverse (NRZI) sequence, such as encoding and storing the first data based on the NRZI encoding algorithm, but is not limited to this; the waveform of the first readback signal is recorded as the sampling waveform of the first readback signal, such as sampling the first readback signal 5 times. Fig.13As shown, there is a small error between the waveforms of the second sequence and the NRZI sequence, but the fourth sequence obtained after sequence detection of the second sequence is basically consistent with the waveform of the NRZI sequence. It can be seen that the method provided in the embodiment of the present application can reduce the influence of nonlinear noise on the first readback signal, and can retain the original characteristics of the first readback signal, and reduce the probability of losing the characteristics of the first readback signal due to noise reduction processing.

[0109] exist Fig.14 In the embodiment, the method for noise reduction based on a linear equalizer is recorded as method 1, the method for noise reduction based only on a neural network model is recorded as method 2, and the method provided in the embodiment of the present application is recorded as method 3. Fig.14 As shown, the detection bit error rate of noise reduction based only on the neural network model is lower than the detection bit error rate of noise reduction based on the linear equalizer, and the detection bit error rate of noise reduction performed by the method provided in the embodiment of the present application is lower than the detection error rate of noise reduction based only on the neural network model.

[0110] It should be noted that, in the case of high-density storage devices with small spacing, multi-track synchronous processing can be adopted, or expanded to two-dimensional readback signals, so as to simultaneously perform noise reduction processing and reduce inter-code crosstalk and inter-track crosstalk. For example, the signal processing device can use the multi-track readback signals as training samples to train the neural network model, and in the subsequent data acquisition process, the multi-track readback signals are input into the trained neural network model. The processing flow and structure of the signal processing device can refer to the description of the aforementioned embodiments, and will not be repeated here.

[0111] The method steps in the embodiments of the present application may be implemented by hardware, or by a processor executing a computer program or instruction. The computer program or instruction may constitute a computer program product.

[0112] The present application also provides a computer program product, which includes computer executable instructions. In one embodiment, the computer executable instructions are used to enable a computer to execute Figure 6 or Fig. 9 Functionality in the method embodiment shown.

[0113] The computer executable instructions may be stored in a computer readable storage medium. The present application also provides a computer readable storage medium, wherein the computer readable storage medium stores the executable instructions. In one embodiment, the computer executable instructions are used to enable the computer to execute Figure 6 or Fig. 9 The functions of any of the method embodiments shown in .

[0114] The present application also provides an electronic device for executing Figure 6 or Fig. 9The method shown.

[0115] The present application also provides a chip for reading a computer program stored in a memory and executing Figure 6 or Fig. 9 The method shown.

[0116] The present application also provides a communication system, which includes a signal processing device and / or a model training device. Figures 2 to 5 The signal processing device in any of the embodiments shown. Please refer to the model training device Figure 8 The model training device in the illustrated embodiment.

[0117] In the embodiments of the present application, "multiple" may refer to two or more than two. In view of this, in the embodiments of the present application, "multiple" may also be understood as "at least two". "At least one" may be understood as one or more, for example, one, two or more. For example, "including at least one" means including one, two or more. For example, including at least one of A, B and C, then A, B, C, A and B, A and C, B and C, or A, B and C may be included. "And / or" describes the association relationship of associated objects. Specifically, there may be three relationships. For example, A and / or B may represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / ", unless otherwise specified, generally indicates that the previously associated objects are in an "or" relationship.

[0118] Unless otherwise specified, ordinal numbers such as "first" and "second" mentioned in the embodiments of the present application are used to distinguish multiple objects, and are not used to limit the order, timing, priority or importance of the multiple objects. For example, the first readback signal and the second readback signal are used to distinguish two readback signals, and do not limit the priority or importance of the two readback signals.

[0119] The computer-readable storage medium provided in the embodiments of the present application may be a random access memory (RAM), a flash memory, a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a register, a hard disk, a mobile hard disk, a CD-ROM, or any other form of computer-readable storage medium known in the art.

[0120] Computer executable instructions may be stored in a computer readable storage medium or transferred from one computer readable storage medium to another computer readable storage medium. For example, the computer program or instructions may be transferred from one website, computer, server or data center to another website, computer, server or data center by wired or wireless means. The computer readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; it may also be an optical medium, such as a digital video disc (DVD); it may also be a semiconductor medium, such as a solid state drive.

[0121] In each embodiment of the present application, if there is no special explanation and logical conflict, the terms and / or descriptions between different embodiments are consistent and can be referenced to each other, and the technical features in different embodiments can be combined to form new embodiments according to their inherent logical relationships. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, including a series of steps or units. Methods, systems, products or devices are not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0122] Although the present application has been described in conjunction with specific features and embodiments thereof, it is apparent that various modifications and combinations may be made thereto without departing from the spirit and scope of the present application. Accordingly, the present specification and the accompanying drawings are merely exemplary illustrations of the schemes defined by the appended claims, and are deemed to have covered any and all modifications, variations, combinations or equivalents within the scope of the present application.

[0123] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A signal processing method, characterized in that: The method comprises: Acquire a first readback signal, where the first readback signal is a sensing signal corresponding to a read operation; Inputting a first sequence and the first readback signal into a first neural network model to obtain a second sequence output by the first neural network model, wherein the first neural network model is used to perform noise reduction processing on the first readback signal, the first sequence is a sequence obtained by performing sequence detection on a third sequence, and the third sequence is a sequence output by the first neural network model based on the first readback signal; performing sequence detection on the second sequence to obtain a fourth sequence; First data corresponding to the first readback signal is acquired according to the fourth sequence.

2. The method according to claim 1, characterized in that The obtaining of the first readback signal comprises: receiving a read instruction, wherein the read instruction is used to read the first data in the first storage device; In response to the read instruction, a read operation is performed on the first storage device to obtain the first read-back signal.

3. The method according to claim 2, characterized in that The first neural network model corresponds to parameters of the first storage device; or, The sample set used to train the first neural network model corresponds to the parameters of the first storage device.

4. The method according to any one of claims 1 to 3, characterized in that The number of the first sequence is one, the number of the third sequence is one, and the method further includes: Before inputting the first sequence and the first readback signal into the first neural network model, inputting the first readback signal into the first neural network model to obtain the third sequence output by the first neural network model; Perform sequence detection on the third sequence to obtain the first sequence.

5. The method according to claim 4, characterized in that The step of inputting the first readback signal into the first neural network model to obtain the third sequence output by the first neural network model includes: The first readback signal and the all-zero sequence are input into the first neural network model to obtain the third sequence.

6. The method according to any one of claims 1 to 3, characterized in that The number of the first sequences is N, the number of the third sequences is N, and N is an integer greater than 1; The step of inputting the first sequence and the first readback signal into a first neural network model to obtain a second sequence output by the first neural network model comprises: Inputting the first readback signal into the first neural network model to obtain a first third sequence among the N third sequences output by the first neural network model; Performing sequence detection on the first third sequence to obtain a first first sequence among the N first sequences; For values ​​of i ranging from 1 to N-1, the following steps are performed: the first readback signal and the i-th first sequence among the N first sequences are input into the first neural network model to obtain the i+1-th third sequence among the third sequences output by the first neural network model; sequence detection is performed on the i+1-th third sequence to obtain the i+1-th first sequence among the N first sequences; wherein i is a positive integer less than N; The first readback signal and the Nth first sequence among the N first sequences are input into the first neural network model to obtain the second sequence.

7. The method according to any one of claims 4 to 6, characterized in that The number of the first sequences is determined during the training process of the first neural network model.

8. The method according to any one of claims 1 to 7, characterized in that A signal quality calculated based on the fourth sequence and the second sequence is greater than or equal to a first threshold.

9. The method according to any one of claims 1 to 8, characterized in that The method further comprises: Acquire M training samples, where the M training samples are sensing signals obtained by performing a read operation on a storage device, and M is a positive integer; For values ​​of j ranging from 1 to M, the following steps are performed to obtain the first neural network model: the fifth sequence and the j-th training sample among the M training samples are input into the second neural network model to obtain the sixth sequence output by the second neural network model, wherein the fifth sequence is a sequence obtained by performing sequence detection on the seventh sequence, and the seventh sequence is a sequence output by the second neural network model based on the j-th training sample; when the sixth sequence meets the training termination condition, the trained second neural network model is obtained; wherein the second neural network model is an initial neural network model or a neural network model obtained by training based on the j-1-th training sample among the M training samples, and j is an integer greater than 0 and less than or equal to M.

10. The method according to claim 9, characterized in that There are multiple storage devices, and parameters of each storage device in the multiple storage devices are the same as parameters of a first storage device, and the first storage device is used to store the first data.

11. The method according to claim 9 or 10, characterized in that: The activation function used by the hidden layer of the second neural network model is a nonlinear activation function.

12. A signal processing device, characterized in that: The device comprises a noise reduction module and a decoding module; wherein: The noise reduction module is used to obtain a first readback signal, where the first readback signal is a sensing signal corresponding to a read operation; input a first sequence and the first readback signal into a first neural network model to obtain a second sequence output by the first neural network model, where the first neural network model is used to perform noise reduction processing on the first readback signal, where the first sequence is a sequence obtained by performing sequence detection on a third sequence, where the third sequence is a sequence output by the first neural network model based on the first readback signal; and perform sequence detection on the second sequence to obtain a fourth sequence; The decoding module is used to obtain first data corresponding to the first readback signal according to the fourth sequence.

13. The device according to claim 12, characterized in that The noise reduction module includes a first submodule and a second submodule, the output of the first submodule is connected to the input of the second submodule, the output of the second submodule is connected to the input of the first submodule, the output of the second submodule is also connected to the decoding module, the number of the first sequence is one, and the number of the third sequence is one; When the first readback signal is input into the first neural network model to obtain the third sequence output by the first neural network model: The first submodule is used to input the first readback signal into the first neural network model to obtain the third sequence output by the first neural network model, and send the third sequence to the second submodule; The second submodule is configured to receive the third sequence, perform sequence detection on the third sequence to obtain the first sequence, and send the first sequence to the first submodule; The first submodule is further used to receive the first sequence, input the first sequence and the first readback signal into the first neural network model, and obtain the second sequence.

14. The device according to claim 12, characterized in that The noise reduction module includes a first submodule and a second submodule, the output of the first submodule is connected to the input of the second submodule, the output of the second submodule is connected to the input of the first submodule, the output of the second submodule is also connected to the decoding module, the number of the first sequences is N, the number of the third sequences is N, and N is an integer greater than 1; When the first sequence and the first readback signal are input into the first neural network model to obtain the second sequence output by the first neural network model: The first submodule is used to input the first readback signal into the first neural network model to obtain a first third sequence among the N third sequences output by the first neural network model, and send the first third sequence to the second submodule; The second submodule is configured to receive the first third sequence, perform sequence detection on the first third sequence to obtain a first first sequence among the N first sequences, and send the first first sequence to the first submodule; For the value of i ranging from 1 to N-1: the first submodule is further used to receive the i-th first sequence among the N first sequences, input the first readback signal and the i-th first sequence into the first neural network model, obtain the i+1-th third sequence among the third sequences output by the first neural network model, and send the i+1-th third sequence to the second submodule; the second submodule is further used to receive the i+1-th third sequence, perform sequence detection on the i+1-th third sequence, obtain the i+1-th first sequence among the N first sequences, and send the i+1-th first sequence to the first submodule; wherein, i is a positive integer less than N; The first submodule is further used to receive the Nth first sequence among the N first sequences, and input the Nth first sequence and the first readback signal into the first neural network model to obtain the second sequence.

15. The device according to any one of claims 12 to 14, characterized in that The device also includes a collection module and a training module; The acquisition module is used to obtain M training samples, where the M training samples are sensing signals obtained by performing a read operation on the storage device, and M is a positive integer; The training module is used to perform the following steps for the value of j from 1 to M to obtain the first neural network model: input the fifth sequence and the j-th training sample among the M training samples into the second neural network model to obtain the sixth sequence output by the second neural network model, wherein the fifth sequence is a sequence obtained by sequence detection of the seventh sequence, and the seventh sequence is a sequence output by the second neural network model based on the j-th training sample; when the sixth sequence meets the training termination condition, the trained second neural network model is obtained; wherein the second neural network model is an initial neural network model or a neural network model obtained by training based on the j-1-th training sample among the M training samples, and j is an integer greater than 0 and less than or equal to M.

16. A signal processing device, characterized in that: The signal processing device comprises an equalizer, a detector and a decoder, wherein the output of the equalizer is connected to the input of the detector, the output of the detector is connected to the input of the equalizer, and the output of the detector is also connected to the decoder, wherein: The equalizer is used to obtain a first readback signal, input the first readback signal into a first neural network model, obtain a third sequence output by the first neural network model, and send the third sequence to the detector, wherein the first readback signal is a sensing signal corresponding to a read operation, and the first neural network model is used to perform noise reduction processing on the first readback signal; The detector is configured to receive the third sequence, perform sequence detection on the third sequence to obtain a first sequence, and send the first sequence to the equalizer; The equalizer is further configured to receive the first sequence, input the first sequence and the first readback signal into the first neural network model, obtain a second sequence output by the first neural network model, and send the second sequence to the detector; The detector is further configured to receive the second sequence, perform sequence detection on the second sequence to obtain a fourth sequence, and send the fourth sequence to the decoder; The decoder is used to obtain first data corresponding to the first readback signal according to the fourth sequence.

17. A signal processing device, characterized in that: The signal processing device comprises an equalizer, a detector and a decoder, wherein the output of the equalizer is connected to the input of the detector, the output of the detector is connected to the input of the equalizer, and the output of the detector is also connected to the decoder, wherein: The equalizer is used to obtain a first readback signal, input the first readback signal into a first neural network model, obtain a first third sequence among N third sequences output by the first neural network model, and send the first third sequence to the detector, wherein the first readback signal is a sensing signal corresponding to a read operation, the first neural network model is used to perform noise reduction processing on the first readback signal, and N is an integer greater than 1; The detector is used to receive the first third sequence, perform sequence detection on the first third sequence to obtain a first first sequence among N first sequences, and send the first first sequence to the equalizer; For the value of i ranging from 1 to N-1: the equalizer is further used to receive the i-th first sequence among the N first sequences, input the first readback signal and the i-th first sequence into the first neural network model, obtain the i+1-th third sequence among the third sequences output by the first neural network model, and send the i+1-th third sequence to the detector; the detector is further used to receive the i+1-th third sequence, perform sequence detection on the i+1-th third sequence, obtain the i+1-th first sequence among the N first sequences, and send the i+1-th first sequence to the equalizer; wherein i is a positive integer less than N; The equalizer is further configured to receive an Nth first sequence among the N first sequences, input the Nth first sequence and the first readback signal into the first neural network model, obtain a second sequence output by the first neural network model, and send the second sequence to the detector; The detector is further configured to receive the second sequence, perform sequence detection on the second sequence to obtain the fourth sequence, and send the fourth sequence to the decoder; The decoder is used to obtain first data corresponding to the first readback signal according to the fourth sequence.

18. A model training method, characterized in that: The method comprises: Acquire a second readback signal, where the second readback signal is a sensing signal obtained by performing a read operation on the storage device; Inputting the fifth sequence and the second readback signal into a second neural network model to obtain a sixth sequence output by the second neural network model, wherein the fifth sequence is a sequence obtained by performing sequence detection on a seventh sequence, and the seventh sequence is a sequence output by the second neural network model based on the second readback signal; When the sixth sequence meets the training termination condition, the trained second neural network model is obtained, and the trained second neural network model is used to perform noise reduction processing on the readback signal.

19. The method according to claim 18, characterized in that The activation function used by the hidden layer of the second neural network model is a nonlinear activation function.

20. A model training device, characterized in that: The device comprises a collection module and a training module, wherein: The acquisition module is used to obtain a second readback signal, where the second readback signal is a sensing signal obtained by performing a read operation on the storage device; The training module is used to input the fifth sequence and the second readback signal into the second neural network model to obtain a sixth sequence output by the second neural network model, wherein the fifth sequence is a sequence obtained by performing sequence detection on the seventh sequence, and the seventh sequence is a sequence output by the second neural network model based on the second readback signal; and when the sixth sequence meets the training termination condition, the trained second neural network model is obtained, and the trained second neural network model is used to perform noise reduction processing on the readback signal.

21. A chip, characterized in that: The chip is used to read a computer program stored in a memory, and execute the method according to any one of claims 1 to 11, or execute the method according to claim 18 or 19.

22. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed on a processor, the processor executes the method according to any one of claims 1 to 11, or executes the method according to claim 18 or 19.

23. A computer program product comprising instructions, characterized in that When the instructions are executed by a processor, the processor is caused to execute the method according to any one of claims 1 to 11, or execute the method according to claim 18 or 19.

Images