Noise suppression and signal identification method based on deep learning

Abstract

The invention discloses a noise suppression and signal identification method based on deep learning, and the method comprises the following steps: dividing a mixed signal containing unstable noise into a plurality of time frames, carrying out the normalization processing, and inputting the time frames into an improved Hebbian neural network. In each frame, a fluctuation characteristic of the signal is calculated, generating a factor for adjusting the learning intensity. The neural network screens out a signal propagation path according to the activation degree in the processing process, adjusts the connection strength based on the correlation between input and output signals, and gradually weakens the weight of the connection which is not activated. And finally, outputting a signal after noise suppression and an identification result. The system is clear in structure, reasonable in processing flow, and suitable for sound signal analysis and processing in a complex environment.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and signal processing technology, and in particular to a noise suppression and signal recognition method based on deep learning. Background Art

[0002] With the widespread adoption of technologies such as voice communication, intelligent perception, and human-computer interaction, signal processing in complex environments has become a key research direction. This is particularly true in scenarios such as traffic monitoring, medical diagnosis, and smart homes, which place higher demands on signal recognition and noise suppression in the presence of non-stationary noise. Traditional signal processing methods, often based on filter design or statistical models, often struggle to balance signal fidelity and noise suppression when faced with dynamic background noise, sudden interference, and nonlinear signal propagation characteristics.

[0003] In the existing technology, although some studies have attempted to introduce neural network methods for noise processing, there are still obvious deficiencies in the following aspects:

[0004] 1. Lack of dynamic modeling mechanism for non-steady-state noise: Existing methods mostly rely on static parameters or fixed structures to process noise and are unable to dynamically adjust the suppression strategy according to signal changes, resulting in limited suppression effect.

[0005] 2. Single neural network structure update mechanism: Traditional Hebbian learning methods often update weights based on fixed linkage rules, ignoring the relationship between the local characteristics of the input signal and the network structure state, which can easily cause network overfitting or invalid connection redundancy problems.

[0006] 3. Lack of effective path screening and learning control mechanisms: Existing neural network methods do not introduce a response strength evaluation mechanism in candidate path selection, resulting in a lack of accuracy in path decision-making. At the same time, the learning intensity cannot be adjusted according to the time-varying characteristics of the signal, making it difficult to adapt to dynamic environments.

[0007] 4. Imperfect long-term update strategy for synaptic connections: Most methods do not consider the timely decay mechanism when synaptic connections are inactive for a long time. In the long run, the network structure has an accumulation of invalid connections, which affects the overall recognition performance and efficiency.

[0008] Therefore, how to provide a noise suppression and signal recognition method based on deep learning is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0009] One purpose of the present invention is to propose a noise suppression and signal recognition method based on deep learning. The present invention adopts an improved Hebbian neural network, combines the generation of dynamic synaptic inhibitory factors, the response intensity-driven path screening mechanism with the weight enhancement and exponential decay mechanism of synaptic connections, and completely constructs a highly adaptable signal processing process under non-steady-state noise interference. It has the advantages of strong adaptability, high recognition accuracy and dynamically adjustable network structure.

[0010] The noise suppression and signal recognition method based on deep learning according to an embodiment of the present invention includes the following steps:

[0011] S1. Acquire an original mixed signal containing non-stationary noise interference, perform time window division and amplitude normalization on the original mixed signal, and form time series frame signal data;

[0012] S2. Calculate the local time fluctuation rate index for each frame of time series frame signal data, and generate a dynamic synaptic inhibition factor based on the local time fluctuation rate index;

[0013] S3, inputting the time series frame signal data into the improved Hebbian neural network, and generating multiple candidate signal propagation paths through neuron activation operations;

[0014] S4. Compare the neuron response strengths of all candidate signal propagation paths and select the signal propagation path with the largest response strength as the only valid signal propagation path in the current iteration cycle based on a winner-takes-all mechanism;

[0015] S5. In the current only valid signal propagation path, based on the correlation between the synaptic input signal and the neuron output signal, perform Hebbian weight reinforcement learning operation under the regulation of dynamic synaptic inhibitory factors;

[0016] S6. Executing a weight decay mechanism based on a combination condition on all synaptic connections in the improved Hebbian neural network, while satisfying a threshold of a time period during which the synaptic connection is continuously inactive and a condition of marking an ineffective signal propagation path, performing a synaptic weight decay operation in the form of an exponential function on the synaptic connections;

[0017] S7. Output the noise suppression result and signal recognition result after being processed by the improved Hebbian neural network.

[0018] Optionally, the time window division and amplitude normalization processing specifically include: using a sliding time window to divide the original mixed signal frame by frame, setting the window length and sliding step parameters so that there is an overlapping interval of a predetermined proportion between adjacent time frame signal segments, and when performing each frame signal extraction, dynamically adjusting the normalized reference interval of the current frame based on the amplitude distribution characteristics of the adjacent previous frame signal segment, within each frame signal segment, setting a linear mapping boundary according to the maximum amplitude and the minimum amplitude, and retaining the amplitude labels of outliers outside the boundary during the normalization process, and continuously splicing all normalized time frame signal data according to the signal sampling timing to generate time series frame signal data.

[0019] Optionally, the S2 specifically includes:

[0020] S21. For each frame of time series frame signal data, set an initial local sampling window, dynamically adjust the length of the local sampling window according to the amplitude change rate between adjacent sampling points in the current time series frame signal data, and extract the amplitude sequence within the adjusted local sampling window;

[0021] S22. Based on the amplitude sequence extracted within the local sampling window, calculate the mean and standard deviation of the amplitude sequence, and use the standard deviation as a local time volatility index corresponding to the corresponding time series frame signal data;

[0022] S23, inputting the local time volatility index into a preset nonlinear inhibition function model, performing an exponential mapping operation, and obtaining a dynamic synaptic inhibition factor corresponding to the time series frame signal data;

[0023] S24. Bind the dynamic synaptic inhibition factor with the corresponding time series frame signal data and output it as a learning strength adjustment parameter for the synaptic connection weight update stage in the improved Hebbian neural network and input it into step S5.

[0024] Optionally, the S3 specifically includes:

[0025] S31, inputting the bound dynamic synaptic inhibitory factor and the corresponding time series frame signal data as input data vectors into an improved Hebbian neural network constructed with a static synaptic connection structure and a dynamic weight update mechanism, so as to activate neuronal responses;

[0026] S32. In the improved Hebbian neural network, for each neuron in the network, calculating the corresponding neuron activation value according to the weighted inner product between the corresponding synaptic connection weight and the input data vector;

[0027] S33, comparing all neuron activation values ​​with a set response threshold, screening out an activation response neuron set whose neuron activation value is greater than or equal to the set response threshold, as the current activation response neuron set;

[0028] S34. Based on the activated response neuron set, a candidate signal propagation path set corresponding to the current time series frame signal data is constructed from all valid synaptic connection paths from the input layer to each neuron in the set, which is used to compare the neuron response intensity and screen the only valid signal propagation path.

[0029] Optionally, the improved Hebbian neural network specifically includes:

[0030] Initialize the synaptic connection weight matrix between neurons, build the static synaptic structure between input units and output neurons, and set the learning rate parameters, neuron activation function, dynamic synaptic inhibition factor sensitivity coefficient and synaptic weight time decay coefficient;

[0031] The bound dynamic synaptic inhibitory factor and the corresponding time series frame signal data are input as input data vectors into the improved Hebbian neural network. The activation values ​​of all output neurons are obtained by calculating the weighted inner product between each output neuron and the input data and applying it to the activation function.

[0032] During the synaptic connection weight update process, for each input unit, the dynamic synaptic inhibition factor α is calculated based on the variance of the amplitude change in the corresponding local time window. i (t):

[0033] α i (t) = exp(-λ·Var(x i (t)));

[0034] Among them, α i (t) represents the dynamic synaptic inhibition factor of input unit i at time t, λ is the synaptic inhibition sensitivity coefficient, Var(x i (t)) is the amplitude variance of input unit i in the local time window, x i (t) is the amplitude of the input unit at time t;

[0035] Introducing the dynamic synaptic inhibition factor into the Hebbian learning rule, the following weight increment expression is formed:

[0036] Δw ij (t) = η·α i (t)·x i (t)·y j (t);

[0037] Where Δwij (t) is the weight increment between input unit i and output neuron j at time t, η is the learning rate, x i (t) is the amplitude of the input unit at time t, y j (t) is the activation value of output neuron j at time t;

[0038] A time-dependent weight decay mechanism is introduced for all synaptic connections. When a synaptic connection does not have a valid activation event within a preset period, an exponential decay operation is triggered and the following weight updates are performed:

[0039] w ij (t+1)=(1-γ·δ ij (t))·w ij (t)+Δw ij (t);

[0040] Among them, w ij (t+1) is the synaptic connection weight at time t+1, γ is the synaptic weight attenuation coefficient, δ ij (t) is the activation monitoring flag variable, indicating whether the synaptic connection is continuously inactive in the current cycle. If the synaptic connection w ij If it is not activated within the set continuous time period, the value is 1, otherwise it is 0. ij (t) is the synaptic weight value at time t.

[0041] After each synaptic weight update, the synaptic connection weight vector is normalized to control the total synaptic connection strength corresponding to each output neuron within a fixed range and maintain the dynamic stability of the network.

[0042] Optionally, the S4 specifically includes:

[0043] S41. For each candidate signal propagation path, extract the neuron activation values ​​corresponding to all activated response neurons in the candidate signal propagation path, and construct a neuron activation value set corresponding to the candidate signal propagation path;

[0044] S42, normalizing the set of neuron activation values ​​corresponding to the candidate signal propagation path, calculating the average activation value and the maximum activation value of the set of neuron activation values ​​corresponding to the candidate signal propagation path, and defining the ratio of the average activation value to the maximum activation value as a path response strength index of the candidate signal propagation path;

[0045] S43, comparing the path response strength indices of all candidate signal propagation paths, and determining the candidate signal propagation path with the largest path response strength index value as the only valid signal propagation path in the current iteration cycle;

[0046] S44. Based on the winner-takes-all mechanism, a freeze marking operation is performed on all candidate signal propagation paths except the only valid signal propagation path. The freeze marking operation is used to prohibit the synaptic connections in the candidate signal propagation paths in the frozen state from participating in the synaptic connection weight update in the current iteration cycle, and only allow the synaptic connections in the only valid signal propagation path to participate in the synaptic connection weight update.

[0047] Optionally, the S5 specifically includes:

[0048] S51. Extracting the synaptic input signal and neuron output signal corresponding to the synaptic connection in the current only valid signal propagation path, where the synaptic input signal is the amplitude of the input unit in the current time frame, and the neuron output signal is the activation response value of the output neuron in the current time frame;

[0049] S52. Determine whether the synaptic connection has a positive response correlation by analyzing the direction of synchronous changes between the synaptic input signal and the neuron output signal. Positive response correlation is a condition in which the synaptic input signal and the neuron output signal have consistent change trends within the same time frame.

[0050] S53. For synaptic connections with positive response correlation, calling a dynamic synaptic inhibitor, the dynamic synaptic inhibitor is used to adjust the learning intensity of the synaptic connection weight enhancement, and the value of the dynamic synaptic inhibitor changes with the local fluctuation of the synaptic input signal;

[0051] S54. Perform a Hebbian weight enhancement learning operation on the synaptic connection that satisfies the positive response correlation and is not in a frozen state in the current only valid signal propagation path. The Hebbian weight enhancement learning operation uses the synaptic input signal, the neuron output signal and the dynamic synaptic inhibition factor as joint input conditions to adjust the current weight value of the synaptic connection, and uses the updated synaptic connection weight value as the initial state of the synaptic connection in the next time frame.

[0052] Optionally, the S6 specifically includes:

[0053] S61. In the improved Hebbian neural network, an activation state tracking structure is established for each synaptic connection to record activation history information of the synaptic connection in a continuous time frame;

[0054] S62. In each iteration cycle, the following determination is performed on each synaptic connection: if the synaptic connection is not activated in all time frames within a preset continuous time period, and the signal propagation path to which the synaptic connection belongs is not marked as the only valid signal propagation path within the corresponding time frame, then the corresponding synaptic connection is determined to meet the weight decay condition;

[0055] S63. For synaptic connections that meet the weight decay condition, perform an exponential function update operation on the current weight value of the synaptic connection, where the base and power exponent parameters of the exponential function are determined by the inactive duration of the synaptic connection and the network state parameters;

[0056] S64. Write the synaptic connection weight value updated in the form of the exponential function into the synaptic connection weight storage structure in the improved Hebbian neural network as the initial synaptic connection weight input for the next time frame.

[0057] Optionally, the S7 specifically includes: performing a data extraction operation on each synaptic connection in the only valid signal propagation path in the current iteration cycle, extracting the synaptic connection weight value corresponding to each synaptic connection and the activation response value of the neuron at the synaptic connection output end, combining the extracted synaptic connection weight value and activation response value in a time series order, constructing a time series reconstructed signal data and performing inter-frame boundary smoothing and amplitude denormalization operations to generate a signal sequence output result after noise suppression, performing a signal category recognition operation based on the distribution of neuron activation response values ​​and synaptic connection structure characteristics in the only valid signal propagation path, generating a corresponding signal identification label, binding the signal sequence output result after noise suppression to the signal identification label at the corresponding frame level, and outputting the time series frame signal processing result data.

[0058] The beneficial effects of the present invention are:

[0059] (1) The present invention realizes the dynamic modeling and recognition of signals in a non-stationary noise interference environment by constructing a complete method system including time series frame signal preprocessing, local time fluctuation rate analysis, improved Hebbian neural network path generation and selection mechanism, weight learning mechanism involving dynamic synaptic inhibitory factors, and exponential weight decay mechanism, thus overcoming the problems of poor adaptability of traditional methods in processing non-stationary signals and rigid network learning mechanism.

[0060] (2) The present invention introduces a winner-takes-all mechanism based on the response strength index, local synchronous synaptic update constraints, and a dynamic synaptic factor adjustment method to achieve high-precision propagation path selection and synaptic plasticity adjustment without increasing network complexity, so that signal recognition can be more focused on key propagation channels, thereby improving the self-organizing ability and selective response ability within the network.

[0061] (3) The present invention combines the activation tracking of synaptic connections in the improved Hebbian neural network with the exponential decay strategy based on non-activation conditions to realize the dynamic elimination mechanism of synaptic weights, improve the adaptive reconstruction capability of the network connection structure, avoid the interference problem caused by the long-term accumulation of redundant connections, and thus improve the signal recognition efficiency and long-term learning stability of the overall system in complex noise environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0063] Figure 1 This is a flowchart of the noise suppression and signal recognition method based on deep learning proposed in the present invention;

[0064] Figure 2 Schematic diagram of the improved Hebbian neural network structure and signal propagation path generation mechanism in the present invention;

[0065] Figure 3 This is a processing flow chart of the local time volatility index calculation and dynamic synaptic inhibition factor generation module in the present invention. DETAILED DESCRIPTION

[0066] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0067] refer to Figure 1-Figure 3 , a noise suppression and signal recognition method based on deep learning, comprising the following steps:

[0068] S1. Acquire an original mixed signal containing non-stationary noise interference, perform time window division and amplitude normalization on the original mixed signal, and form time series frame signal data;

[0069] S2. Calculate the local time fluctuation rate index for each frame of time series frame signal data, and generate a dynamic synaptic inhibition factor based on the local time fluctuation rate index;

[0070] S3, inputting the time series frame signal data into the improved Hebbian neural network, and generating multiple candidate signal propagation paths through neuron activation operations;

[0071] S4. Compare the neuron response strengths of all candidate signal propagation paths and select the signal propagation path with the largest response strength as the only valid signal propagation path in the current iteration cycle based on a winner-takes-all mechanism;

[0072] S5. In the current only valid signal propagation path, based on the correlation between the synaptic input signal and the neuron output signal, perform Hebbian weight reinforcement learning operation under the regulation of dynamic synaptic inhibitory factors;

[0073] S6. Executing a weight decay mechanism based on a combination condition on all synaptic connections in the improved Hebbian neural network, while satisfying a threshold of a time period during which the synaptic connection is continuously inactive and a condition of marking an ineffective signal propagation path, performing a synaptic weight decay operation in the form of an exponential function on the synaptic connections;

[0074] S7. Output the noise suppression result and signal recognition result after being processed by the improved Hebbian neural network.

[0075] In this embodiment, the time window division and amplitude normalization processing specifically include: using a sliding time window to divide the original mixed signal frame by frame, setting the window length and sliding step parameters so that there is an overlapping interval of a predetermined proportion between adjacent time frame signal segments, and when performing each frame signal extraction, dynamically adjusting the normalized reference interval of the current frame based on the amplitude distribution characteristics of the adjacent previous frame signal segment, within each frame signal segment, setting a linear mapping boundary according to the maximum amplitude and the minimum amplitude, and retaining the amplitude labels of outliers outside the boundary during the normalization process, and continuously splicing all normalized time frame signal data according to the signal sampling timing to generate time series frame signal data.

[0076] This implementation method divides the original mixed signal frame by frame by setting the sliding time window parameters and introduces an inter-frame overlapping mechanism, which effectively improves the continuous perception capability of non-stationary noise signals. When normalizing each frame signal, it dynamically refers to the amplitude distribution of the previous frame to achieve cross-frame amplitude stability adjustment, and retains the boundary outlier labels on the basis of linear mapping to ensure that key signal features are not compressed or lost. After normalization, a unified time series frame signal data is formed by splicing the signal timing, providing a data input basis with consistent structure and balanced amplitude for subsequent local fluctuation analysis and neural network input, which helps to improve the recognition capability and robustness of the entire system for weak signals.

[0077] In this embodiment, S2 specifically includes:

[0078] S21. For each frame of time series frame signal data, set an initial local sampling window, dynamically adjust the length of the local sampling window according to the amplitude change rate between adjacent sampling points in the current time series frame signal data, and extract the amplitude sequence within the adjusted local sampling window;

[0079] S22. Based on the amplitude sequence extracted within the local sampling window, calculate the mean and standard deviation of the amplitude sequence, and use the standard deviation as a local time volatility index corresponding to the corresponding time series frame signal data;

[0080] S23, inputting the local time volatility index into a preset nonlinear inhibition function model, performing an exponential mapping operation, and obtaining a dynamic synaptic inhibition factor corresponding to the time series frame signal data;

[0081] S24. Bind the dynamic synaptic inhibition factor with the corresponding time series frame signal data and output it as a learning strength adjustment parameter for the synaptic connection weight update stage in the improved Hebbian neural network and input it into step S5.

[0082] This embodiment dynamically adjusts the window length by setting a local sampling window and combining it with the amplitude change rate between adjacent sampling points in the time series frame signal data, thereby obtaining an amplitude sequence with higher time-varying sensitivity, and measuring the local time volatility index with the standard deviation of the amplitude sequence. The volatility index is then mapped into a dynamic synaptic inhibition factor through a nonlinear inhibition function model, and bound to the time series frame signal data for output, thereby realizing a dynamic adjustment mechanism for synaptic learning strength. While improving the ability to recognize non-stationary features in a noisy background, it also enhances the stability and adaptability of the neural network synaptic weight update, effectively enhancing the system's perception and response capabilities to subtle changes in complex signals.

[0083] In this embodiment, S3 specifically includes:

[0084] S31, inputting the bound dynamic synaptic inhibitory factor and the corresponding time series frame signal data as input data vectors into an improved Hebbian neural network constructed with a static synaptic connection structure and a dynamic weight update mechanism, so as to activate neuronal responses;

[0085] S32. In the improved Hebbian neural network, for each neuron in the network, calculating the corresponding neuron activation value according to the weighted inner product between the corresponding synaptic connection weight and the input data vector;

[0086] S33, comparing all neuron activation values ​​with a set response threshold, screening out an activation response neuron set whose neuron activation value is greater than or equal to the set response threshold, as the current activation response neuron set;

[0087] S34. Based on the activated response neuron set, a candidate signal propagation path set corresponding to the current time series frame signal data is constructed from all valid synaptic connection paths from the input layer to each neuron in the set, which is used to compare the neuron response intensity and screen the only valid signal propagation path.

[0088] This implementation combines dynamic synaptic inhibition factors with time-series frame signal data and feeds them into a modified Hebbian neural network with a static synaptic connection structure and a dynamic weight update mechanism. It then uses a weighted inner product approach to calculate neuron activation values ​​one by one, then filters out the set of currently activated response neurons by setting a threshold. It then constructs a set of candidate signal propagation paths from the synaptic connection paths corresponding to the activated neurons, effectively ensuring the network structure's selectivity and sensitivity to input signals. This approach enables the precise extraction of significant feature paths in non-steady-state signal environments, providing a stable and highly responsive candidate foundation for subsequent signal propagation path screening and synaptic learning operations, thereby improving the neural network's perception of target signals and its dynamic recognition efficiency.

[0089] In this embodiment, the improved Hebbian neural network specifically includes:

[0090] Initialize the synaptic connection weight matrix between neurons, build the static synaptic structure between input units and output neurons, and set the learning rate parameters, neuron activation function, dynamic synaptic inhibition factor sensitivity coefficient and synaptic weight time decay coefficient;

[0091] The bound dynamic synaptic inhibitory factor and the corresponding time series frame signal data are input as input data vectors into the improved Hebbian neural network. The activation values ​​of all output neurons are obtained by calculating the weighted inner product between each output neuron and the input data and applying it to the activation function.

[0092] During the synaptic connection weight update process, for each input unit, the dynamic synaptic inhibition factor α is calculated based on the variance of the amplitude change in the corresponding local time window. i (t):

[0093] α i (t) = exp(-λ·Var(x i (t)));

[0094] Among them, α i (t) represents the dynamic synaptic inhibition factor of input unit i at time t, λ is the synaptic inhibition sensitivity coefficient, Var(x i (t)) is the amplitude variance of input unit i in the local time window, x i (t) is the amplitude of the input unit at time t;

[0095] This formula reflects the short-term stability of the input signal by constructing an index of the variance of the input unit's amplitude change within a local time window. It then generates a dynamic synaptic inhibition factor based on an exponential mapping mechanism. This factor exhibits a nonlinear decay with increasing signal volatility, dynamically suppressing unstable input signals during the synaptic connection weight update phase. This mechanism achieves weight adjustment sensitivity control based on the short-term variability of the input signal, enhancing the network's robustness to sudden noise.

[0096] Introducing the dynamic synaptic inhibition factor into the Hebbian learning rule, the following weight increment expression is formed:

[0097] Δw ij (t) = η·α i (t)·x i (t)·y j (t);

[0098] Where Δw ij (t) is the weight increment between input unit i and output neuron j at time t, η is the learning rate, x i (t) is the amplitude of the input unit at time t, y j (t) is the activation value of output neuron j at time t;

[0099] This formula introduces a dynamic synaptic inhibition factor based on the traditional Hebbian weight enhancement rule. The magnitude of the synaptic connection weight enhancement is determined by the product of the input unit amplitude, the output neuron activation response value, and the learning rate. The introduction of the dynamic synaptic inhibition factor dynamically adjusts the learning intensity, explicitly suppressing the impact of highly volatile input signals on synaptic weights. This improves the network's ability to adapt to transient abnormal fluctuations and effectively prevents synaptic connections from being over-strengthened due to unstable signals.

[0100] A time-dependent weight decay mechanism is introduced for all synaptic connections. When a synaptic connection does not have a valid activation event within a preset period, an exponential decay operation is triggered and the following weight updates are performed:

[0101] w ij (t+1)=(1-γ·δ ij (t))·w ij (t)+Δw ij (t);

[0102] Among them, w ij (t+1) is the synaptic connection weight at time t+1, γ is the synaptic weight attenuation coefficient, δ ij (t) is the activation monitoring flag variable, if the synaptic connection w ijIf it is not activated within the set continuous time period, the value is 1, otherwise it is 0. ij (t) is the synaptic weight value at time t.

[0103] This formula establishes a time-dependent weight decay mechanism that automatically adjusts the weight of synaptic connections when they remain inactive for extended periods. By setting an activation monitoring flag variable, it determines whether a synaptic connection has been effectively activated within a preset period. If it remains inactive for extended periods, an exponential decay process is triggered. This mechanism applies an exponential decay factor to synaptic connections that remain inactive for extended periods, thereby dynamically eliminating ineffective synaptic connections and helping to maintain the compactness of the network structure and the stability of the weight distribution.

[0104] After each synaptic weight update, the synaptic connection weight vector is normalized to control the total synaptic connection strength corresponding to each output neuron within a fixed range and maintain the dynamic stability of the network.

[0105] This implementation initializes the static synaptic structure and related parameters in a modified Hebbian neural network, uses time series frame signal data and bound dynamic synaptic inhibitors as input, calculates the output neuron activation value and drives signal propagation, introduces a dynamic synaptic inhibitor based on the local volatility of the input signal to participate in the weight learning strength adjustment, and constructs a time-dependent weight decay mechanism, which automatically causes synaptic connections to decay exponentially under long-term inactivity conditions. After each round of weight update, normalization is performed to maintain the total synaptic strength of the output neurons stable. This implementation effectively achieves dynamic plasticity and stability control of neural network weight updates, improving the responsiveness and robustness of noise suppression and signal recognition in non-steady-state signal environments.

[0106] In this embodiment, the S4 specifically includes:

[0107] S41. For each candidate signal propagation path, extract the neuron activation values ​​corresponding to all activated response neurons in the candidate signal propagation path, and construct a neuron activation value set corresponding to the candidate signal propagation path;

[0108] S42, normalizing the set of neuron activation values ​​corresponding to the candidate signal propagation path, calculating the average activation value and the maximum activation value of the set of neuron activation values ​​corresponding to the candidate signal propagation path, and defining the ratio of the average activation value to the maximum activation value as a path response strength index of the candidate signal propagation path;

[0109] S43, comparing the path response strength indices of all candidate signal propagation paths, and determining the candidate signal propagation path with the largest path response strength index value as the only valid signal propagation path in the current iteration cycle;

[0110] S44. Based on the winner-takes-all mechanism, a freeze marking operation is performed on all candidate signal propagation paths except the only valid signal propagation path. The freeze marking operation is used to prohibit the synaptic connections in the candidate signal propagation paths in the frozen state from participating in the synaptic connection weight update in the current iteration cycle, and only allow the synaptic connections in the only valid signal propagation path to participate in the synaptic connection weight update.

[0111] This implementation extracts the activation values ​​of all activated responsive neurons in each candidate signal propagation path, constructs a set of neuron activation values, and calculates the ratio of the average activation value to the maximum activation value after normalization as a path response strength indicator. This then performs a strength comparison across multiple candidate paths to determine a single valid signal propagation path. A winner-takes-all mechanism is then used to freeze the synaptic connections of unselected paths, preventing them from participating in weight updates during the current iteration. This mechanism effectively strengthens the signal expression of the optimal path, eliminates learning interference from interfering paths, and enables the network structure to perform targeted reinforcement learning on highly responsive paths, thereby improving signal discrimination accuracy and model focusing capabilities in noisy environments.

[0112] In this embodiment, the S5 specifically includes:

[0113] S51. Extracting the synaptic input signal and neuron output signal corresponding to the synaptic connection in the current only valid signal propagation path, where the synaptic input signal is the amplitude of the input unit in the current time frame, and the neuron output signal is the activation response value of the output neuron in the current time frame;

[0114] S52. Determine whether the synaptic connection has a positive response correlation by analyzing the direction of synchronous changes between the synaptic input signal and the neuron output signal. Positive response correlation is a condition in which the synaptic input signal and the neuron output signal have consistent change trends within the same time frame.

[0115] S53. For synaptic connections with positive response correlation, calling a dynamic synaptic inhibitor, the dynamic synaptic inhibitor is used to adjust the learning intensity of the synaptic connection weight enhancement, and the value of the dynamic synaptic inhibitor changes with the local fluctuation of the synaptic input signal;

[0116] S54. Perform a Hebbian weight enhancement learning operation on the synaptic connection that satisfies the positive response correlation and is not in a frozen state in the current only valid signal propagation path. The Hebbian weight enhancement learning operation uses the synaptic input signal, the neuron output signal and the dynamic synaptic inhibition factor as joint input conditions to adjust the current weight value of the synaptic connection, and uses the updated synaptic connection weight value as the initial state of the synaptic connection in the next time frame.

[0117] This implementation extracts synaptic input signals and neuronal output signals from a single valid signal propagation path, analyzes their synchronous variation trends to determine the relevance of positive responses, and introduces a dynamic synaptic inhibition factor associated with the local fluctuations of the input signal to adjust the intensity of Hebbian weight reinforcement learning, thereby dynamically updating synaptic connection weights. While ensuring the validity of the signal propagation path, this mechanism improves the ability and adaptability of the synaptic connection learning process to recognize weakly correlated signals, thereby enhancing the learning efficiency and generalization ability of the neural network for target signals in non-stationary noise, providing a foundation for achieving higher-precision noise suppression and signal recognition.

[0118] In this embodiment, S6 specifically includes:

[0119] S61. In the improved Hebbian neural network, an activation state tracking structure is established for each synaptic connection to record activation history information of the synaptic connection in a continuous time frame;

[0120] S62. In each iteration cycle, the following determination is performed on each synaptic connection: if the synaptic connection is not activated in all time frames within a preset continuous time period, and the signal propagation path to which the synaptic connection belongs is not marked as the only valid signal propagation path within the corresponding time frame, then the corresponding synaptic connection is determined to meet the weight decay condition;

[0121] S63. For synaptic connections that meet the weight decay condition, perform an exponential function update operation on the current weight value of the synaptic connection, where the base and power exponent parameters of the exponential function are determined by the inactive duration of the synaptic connection and the network state parameters;

[0122] S64. Write the synaptic connection weight value updated in the form of the exponential function into the synaptic connection weight storage structure in the improved Hebbian neural network as the initial synaptic connection weight input for the next time frame.

[0123] This implementation establishes an activation state tracking structure for each synaptic connection in a modified Hebbian neural network, recording the activation history of the synaptic connection within continuous time frames. Based on the labeling results of synaptic connections that have not been activated and have not participated in a unique valid signal propagation path for multiple consecutive time periods, it is determined that the connection meets the weight decay condition. Then, a weight update operation in the form of an exponential function is performed on the synaptic connections that meet the conditions. The base and power exponent parameters of the exponential function are jointly determined by the inactivity duration of the synaptic connection and the network state parameters. Finally, the updated weight value is written into the network as the initial input for the next time frame. This method can dynamically adjust the weight values ​​of long-term inactive synaptic connections, prevent redundant connections from interfering with the network learning process, enhance the self-regulation ability of the network structure, and effectively optimize the overall performance of noise suppression and signal recognition.

[0124] In this embodiment, S7 specifically includes: performing a data extraction operation on each synaptic connection in the only valid signal propagation path in the current iteration cycle, extracting the synaptic connection weight value corresponding to each synaptic connection and the activation response value of the neuron at the output end of the synaptic connection, combining the extracted synaptic connection weight value and the activation response value in a time series order, constructing a time series reconstructed signal data and performing inter-frame boundary smoothing and amplitude denormalization operations to generate a signal sequence output result after noise suppression, performing a signal category recognition operation based on the distribution of neuron activation response values ​​and synaptic connection structure characteristics in the only valid signal propagation path, generating a corresponding signal identification label, binding the signal sequence output result after noise suppression to the signal identification label at the corresponding frame level, and outputting the time series frame signal processing result data.

[0125] This embodiment extracts and combines the synaptic connection weight values ​​of each synaptic connection in the only valid signal propagation path and the activation response values ​​of the neurons at the synaptic connection output end to construct a complete time series reconstructed signal data, and generates a noise-suppressed signal output result through inter-frame boundary smoothing and amplitude denormalization operations. Combined with the temporal distribution of the neuron activation response value and the structural characteristics of the synaptic connection, the signal category recognition and label generation are completed, and the signal recognition label is bound to the output result at the corresponding frame level to achieve accurate recognition of the target signal and effective suppression of interference noise, providing structured and semantically clear output results for subsequent processing, and improving the integration and applicability of the overall signal processing link.

[0126] Example 1:

[0127] To verify the feasibility of this invention, it was applied to the processing chip in the voice signal processing module of the dispatch center. The sampling frequency was set to 16 kHz, and each transmitted 10-second raw mixed voice signal stream was partitioned into sliding windows, with each frame containing 400 samples and a frame shift of 160 samples. The raw signal was first normalized to its maximum and minimum amplitudes within each frame, and the normalization reference interval was dynamically adjusted to form continuous time series frame signal data.

[0128] During each frame processing phase, the system dynamically sets the local sampling window size based on the amplitude change rate of the current frame. It also calculates the standard deviation of the amplitude within the window as a volatility indicator. An exponential mapping function model is used to generate a dynamic synaptic inhibition factor, which is bound to each frame of signal data and fed as an input vector into the initialized modified Hebbian neural network. The network structure consists of 128 input units and 64 output neurons. The synaptic connection weight matrix is ​​initially randomly distributed between [0.1 and 0.3], and the activation function is set to the Sigmoid function.

[0129] During the neuron activation phase, the weighted sum of the response values ​​of all neurons is calculated and compared with a set threshold to select a set of activated and responsive neurons. This set is then used to infer the signal propagation path. Among all candidate paths, the ratio of the maximum to the average path activation value is calculated as the path response strength indicator. Based on a winner-takes-all strategy, the path with the strongest response strength is marked as the only valid signal propagation path. The remaining paths are marked as frozen and prohibited from participating in subsequent synaptic weight updates.

[0130] Within a single valid signal propagation path, the input unit value and output neuron activation value of each synaptic connection are extracted to determine whether there is a positive response correlation. For synaptic connections that meet the conditions and are not frozen, the dynamic synaptic inhibition factor generated by the fluctuation rate of the synaptic input signal is used to dynamically adjust the Hebbian learning intensity, performing weight-reinforced learning. After the weight adjustment, the updated value is written to the synaptic connection storage structure as the initial state for the next frame. To prevent the long-term existence of redundant synapses and the resulting network imbalance, the system introduces a weight decay mechanism based on the combined judgment of inactivation history and path validity. Synaptic connections that have been inactive for multiple consecutive time periods and are not part of a valid propagation path are subjected to an exponential weight decay operation, and the synaptic weights are normalized after each update.

[0131] The system combines the weights and neuron activation values ​​of all synaptic connections in a unique valid signal propagation path in time series, performs boundary smoothing and denormalization operations, and generates a noise-suppressed signal output result. It also identifies the signal type based on the activation value distribution and structural characteristics, and outputs labeled time series frame signal data. A field deployment test was conducted at a dispatching center, collecting 10 sets of real train communication recording samples for comparative evaluation. The test results are shown in the following table:

[0132] Table 1: Comparison of signal processing performance based on improved Hebbian neural network

[0133]

[0134] The experimental data results in Table 1 show that, in the real voice communication environment of the Beijing Subway Line 15 dispatching center, the noise suppression and signal recognition method based on the improved Hebbian neural network proposed in this invention achieved an average improvement of approximately 12.9 percentage points in recognition accuracy and nearly 7dB in signal-to-noise ratio compared to the traditional LSTM model. It demonstrated stable robustness and recognition capabilities in complex environments with a high noise content (average 47.8%), and the system processing latency remained consistently within 25 milliseconds, demonstrating good real-time performance. Overall, these results demonstrate that the proposed method can achieve more accurate signal recognition and efficient noise suppression in non-steady-state speech noise environments, making it suitable for voice communication scenarios with high requirements for both accuracy and response speed.

[0135] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A noise suppression and signal recognition method based on deep learning, characterized in that: The following steps are involved: S1. Acquire an original mixed signal containing non-stationary noise interference, perform time window division and amplitude normalization on the original mixed signal, and form time series frame signal data; S2. Calculate the local time fluctuation rate index for each frame of time series frame signal data, and generate a dynamic synaptic inhibition factor based on the local time fluctuation rate index; S3, inputting the time series frame signal data into the improved Hebbian neural network, and generating multiple candidate signal propagation paths through neuron activation operations; S4. Compare the neuron response strengths of all candidate signal propagation paths and select the signal propagation path with the largest response strength as the only valid signal propagation path in the current iteration cycle based on a winner-takes-all mechanism; S5. In the current only valid signal propagation path, based on the correlation between the synaptic input signal and the neuron output signal, perform Hebbian weight reinforcement learning operation under the regulation of dynamic synaptic inhibitory factors; S6. Executing a weight decay mechanism based on a combination condition on all synaptic connections in the improved Hebbian neural network, while satisfying a threshold of a time period during which the synaptic connection is continuously inactive and a condition of marking an ineffective signal propagation path, performing a synaptic weight decay operation in the form of an exponential function on the synaptic connections; S7. Output the noise suppression result and signal recognition result after being processed by the improved Hebbian neural network.

2. The noise suppression and signal recognition method based on deep learning according to claim 1, characterized in that: The time window division and amplitude normalization processing specifically includes: using a sliding time window to divide the original mixed signal frame by frame, setting the window length and sliding step parameters so that there is an overlapping interval of a predetermined proportion between adjacent time frame signal segments, and when performing each frame signal extraction, dynamically adjusting the normalized reference interval of the current frame based on the amplitude distribution characteristics of the adjacent previous frame signal segment, within each frame signal segment, setting a linear mapping boundary according to the maximum amplitude and the minimum amplitude, and retaining the amplitude labels of outliers outside the boundary during the normalization process, and continuously splicing all normalized time frame signal data according to the signal sampling timing to generate time series frame signal data.

3. The noise suppression and signal recognition method based on deep learning according to claim 1, characterized in that: The S2 specifically includes: S21. For each frame of time series frame signal data, set an initial local sampling window, dynamically adjust the length of the local sampling window according to the amplitude change rate between adjacent sampling points in the current time series frame signal data, and extract the amplitude sequence within the adjusted local sampling window; S22. Based on the amplitude sequence extracted within the local sampling window, calculate the mean and standard deviation of the amplitude sequence, and use the standard deviation as a local time volatility index corresponding to the corresponding time series frame signal data; S23, inputting the local time volatility index into a preset nonlinear inhibition function model, performing an exponential mapping operation, and obtaining a dynamic synaptic inhibition factor corresponding to the time series frame signal data; S24. Bind the dynamic synaptic inhibitory factor and the corresponding time series frame signal data and output them.

4. The noise suppression and signal recognition method based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31, inputting the bound dynamic synaptic inhibitory factor and the corresponding time series frame signal data as input data vectors into an improved Hebbian neural network constructed with a static synaptic connection structure and a dynamic weight update mechanism; S32. In the improved Hebbian neural network, for each neuron in the network, calculating the corresponding neuron activation value according to the weighted inner product between the corresponding synaptic connection weight and the input data vector; S33, comparing all neuron activation values ​​with a set response threshold, and screening out an activation response neuron set whose neuron activation value is greater than or equal to the set response threshold; S34. Based on the activated response neuron set, a candidate signal propagation path set corresponding to the current time series frame signal data is constructed from all valid synaptic connection paths from the input layer to each neuron in the set.

5. The noise suppression and signal recognition method based on deep learning according to claim 1, characterized in that: The improved Hebbian neural network specifically includes: Initialize the synaptic connection weight matrix between neurons, build the static synaptic connection structure between input units and output neurons, and set the learning rate parameters, neuron activation function, dynamic synaptic inhibition factor sensitivity coefficient and synaptic weight time decay coefficient; The bound dynamic synaptic inhibitory factor and the corresponding time series frame signal data are input as input data vectors into the improved Hebbian neural network. The activation values ​​of all output neurons are obtained by calculating the weighted inner product between each output neuron and the input data and applying it to the activation function. During the synaptic connection weight update process, for each input unit, the dynamic synaptic inhibition factor α is calculated based on the variance of the amplitude change in the corresponding local time window. i (t): α i (t)=exp(-λ·Var(x i (t))); Among them, α i (t) represents the dynamic synaptic inhibition factor of input unit i at time t, λ is the synaptic inhibition sensitivity coefficient, Var(x i (t)) is the amplitude variance of input unit i in the local time window, x i (t) is the amplitude of the input unit at time t; Introducing the dynamic synaptic inhibition factor into the Hebbian learning rule, the following weight increment expression is formed: Δw ij (t)=η·α i (t)·x i (t)·y j (t); Where Δw ij (t) is the weight increment between input unit i and output neuron j at time t, η is the learning rate, x i (t) is the amplitude of the input unit at time t, y j (t) is the activation value of output neuron j at time t; A time-dependent weight decay mechanism is introduced for all synaptic connections. When a synaptic connection does not have a valid activation event within a preset period, an exponential decay operation is triggered and the following weight updates are performed: w ij (t+1)=(1-γ·δ ij (t))·w ij (t)+Δw ij (t); Among them, w ij (t+1) is the synaptic connection weight at time t+1, γ is the synaptic weight attenuation coefficient, δ ij (t) is the activation monitoring flag variable, if the synaptic connection w ij If it is not activated within the set continuous time period, the value is 1, otherwise it is 0. ij (t) is the synaptic weight value at time t. After each synaptic weight update, the synaptic connection weight vector is normalized.

6. The noise suppression and signal recognition method based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41. For each candidate signal propagation path, extract the neuron activation values ​​corresponding to all activated response neurons in the candidate signal propagation path, and construct a neuron activation value set corresponding to the candidate signal propagation path; S42, normalizing the set of neuron activation values ​​corresponding to the candidate signal propagation path, calculating the average activation value and the maximum activation value of the set of neuron activation values ​​corresponding to the candidate signal propagation path, and defining the ratio of the average activation value to the maximum activation value as a path response strength index of the candidate signal propagation path; S43, comparing the path response strength indices of all candidate signal propagation paths, and determining the candidate signal propagation path with the largest path response strength index value as the only valid signal propagation path in the current iteration cycle; S44. Based on the winner-takes-all mechanism, a freeze marking operation is performed on all candidate signal propagation paths except the only valid signal propagation path. The freeze marking operation is used to prohibit the synaptic connections in the candidate signal propagation paths in the frozen state from participating in the synaptic connection weight update in the current iteration cycle, and only allow the synaptic connections in the only valid signal propagation path to participate in the synaptic connection weight update.

7. The noise suppression and signal recognition method based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51. Extracting the synaptic input signal and neuron output signal corresponding to the synaptic connection in the current only valid signal propagation path, where the synaptic input signal is the amplitude of the input unit in the current time frame, and the neuron output signal is the activation response value of the output neuron in the current time frame; S52. Determine whether the synaptic connection has a positive response correlation by analyzing the direction of synchronous changes between the synaptic input signal and the neuron output signal. Positive response correlation is a condition in which the synaptic input signal and the neuron output signal have consistent change trends within the same time frame. S53. For synaptic connections with positive response correlation, calling a dynamic synaptic inhibitor, the dynamic synaptic inhibitor is used to adjust the learning intensity of the synaptic connection weight enhancement, and the value of the dynamic synaptic inhibitor changes with the local fluctuation of the synaptic input signal; S54. Perform a Hebbian weight enhancement learning operation on the synaptic connection that satisfies the positive response correlation and is not in a frozen state in the current only valid signal propagation path. The Hebbian weight enhancement learning operation uses the synaptic input signal, the neuron output signal and the dynamic synaptic inhibition factor as joint input conditions to adjust the current weight value of the synaptic connection, and uses the updated synaptic connection weight value as the initial state of the synaptic connection in the next time frame.

8. The noise suppression and signal recognition method based on deep learning according to claim 1, characterized in that: The S6 specifically includes: S61. In the improved Hebbian neural network, an activation state tracking structure is established for each synaptic connection to record activation history information of the synaptic connection in a continuous time frame; S62. In each iteration cycle, the following determination is performed on each synaptic connection: if the synaptic connection is not activated in all time frames within a preset continuous time period, and the signal propagation path to which the synaptic connection belongs is not marked as the only valid signal propagation path within the corresponding time frame, then the corresponding synaptic connection is determined to meet the weight decay condition; S63. For synaptic connections that meet the weight decay condition, perform an exponential function update operation on the current weight value of the synaptic connection, where the base and power exponent parameters of the exponential function are determined by the inactive duration of the synaptic connection and the network state parameters; S64. Write the synaptic connection weight value updated in the form of an exponential function into the synaptic connection weight storage structure in the improved Hebbian neural network.

9. The noise suppression and signal recognition method based on deep learning according to claim 1, characterized in that: The S7 specifically includes: performing a data extraction operation on each synaptic connection in the only valid signal propagation path in the current iteration cycle, extracting the synaptic connection weight value corresponding to each synaptic connection and the activation response value of the neuron at the output end of the synaptic connection, combining the extracted synaptic connection weight value and the activation response value in a time series order, constructing a time series reconstructed signal data and performing inter-frame boundary smoothing and amplitude denormalization operations to generate a signal sequence output result after noise suppression, performing a signal category recognition operation based on the distribution of neuron activation response values ​​and synaptic connection structural characteristics in the only valid signal propagation path, generating a corresponding signal identification label, binding the signal sequence output result after noise suppression to the signal identification label at the corresponding frame level, and outputting the time series frame signal processing result data.

Images