Intelligent meter control method and system based on deep learning

Abstract

The application provides a kind of intelligent meter control method and system based on deep learning, it is related to deep learning technical field, first, the original measurement signal data of current measurement period and the historical measurement signal data of last control period of intelligent meter are acquired, the joint decomposition processing of time-frequency domain is respectively executed to two, current measurement signal decomposition features and historical measurement signal decomposition features containing time domain component sequence and frequency domain component sequence are generated.The above-mentioned features are jointly input into the deep time sequence correlation reinforcement network constructed in advance, and the joint inference operation is carried out through the cross-period dependence extraction subnet and signal component interaction subnet in it, and the control parameter adjustment scheme is generated.The instrument control instruction set containing actuator control voltage instruction and signal acquisition module parameter instruction is generated according to the control parameter adjustment scheme, and is sent to actuator and signal acquisition module through control interface.The application effectively improves the precision and self-adaptive ability of intelligent meter control.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a deep learning-based intelligent instrument control method and system. Background Technology

[0002] Intelligent instruments are widely used in industrial process control, energy metering, and environmental monitoring. Their core function is to precisely regulate actuators and optimize signal acquisition parameters based on real-time measurement signals to ensure system stability and measurement accuracy. However, with the increasing complexity of industrial scenarios, the measurement signals acquired by intelligent instruments often exhibit characteristics such as high nonlinearity, multi-frequency superposition, and noise interference, posing significant challenges to the real-time performance and adaptability of control methods.

[0003] In existing technologies, the control methods for intelligent instruments mostly employ classical control strategies based on proportional-integral-derivative (PID) controllers or simple neural network control strategies based on single-cycle data. Classical control strategies rely on manually tuned fixed parameters, making them difficult to adaptively adjust to complex dynamic measurement signals, resulting in insufficient control accuracy and robustness. Neural network control strategies based on single-cycle data typically only extract and map features from the current moment's raw signal, ignoring the temporal evolution and cross-cycle dependencies of the measurement signals between adjacent control cycles. This leads to a lack of consideration for long-term signal trends in the control parameter adjustment scheme. Furthermore, existing technologies mostly analyze measurement signals directly in a single dimension, either the time or frequency domain, failing to fully exploit the complementary features of the signal in both dimensions. This limits the completeness of feature representation and consequently affects the accuracy of control decisions. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a deep learning-based intelligent instrument control method, the method comprising: Acquire the raw measurement signal data collected by the smart instrument during the current measurement cycle and the historical measurement signal data corresponding to the previous control cycle stored in the smart instrument; The original measurement signal data is subjected to joint time-frequency domain decomposition processing to generate measurement signal decomposition features containing time-domain component sequences and frequency-domain component sequences. The historical measurement signal data is subjected to joint time-frequency domain decomposition processing to generate historical measurement signal decomposition features, which contain historical time-domain component sequences and historical frequency-domain component sequences. The measured signal decomposition features and the historical measured signal decomposition features are jointly input into a pre-constructed deep temporal correlation enhancement network. Joint inference operations are performed through the cross-period dependency extraction subnet and signal component interaction subnet within the deep temporal correlation enhancement network to generate a control parameter adjustment scheme. A set of instrument control commands is generated according to the control parameter adjustment scheme, and the set of instrument control commands is sent to the actuator and signal acquisition module through the control interface of the smart instrument. The set of instrument control commands includes actuator control voltage commands and signal acquisition module parameter commands.

[0005] Furthermore, embodiments of the present invention also provide a deep learning-based intelligent instrument control system, comprising: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned deep learning-based intelligent instrument control method by executing the machine-executable instructions.

[0006] Based on the above, by performing joint time-frequency domain decomposition processing on the original measurement signal data of the current measurement cycle and the historical measurement signal data of the previous control cycle, the features of the measurement signal in both the time and frequency domains are extracted and structurally represented simultaneously, overcoming the information loss caused by single-dimensional analysis. On this basis, the decomposition features of the measurement signal of the current cycle and the decomposition features of the measurement signal of the historical cycle are jointly input into a deep temporal correlation reinforcement network. The cross-cycle dependency extraction subnet within this deep temporal correlation reinforcement network is used to mine the temporal evolution law and long-term dependency relationship of signal features between adjacent control cycles. At the same time, the signal component interaction subnet is used to realize deep interactive reasoning between the time domain components and frequency domain components of the current cycle and the time domain components and frequency domain components of the historical cycle. This makes the generated control parameter adjustment scheme fully consider the cross-cycle evolution trend of the signal and integrate the interactive feature information of multiple time domains and multiple frequency domains, thereby significantly improving the control parameter adjustment accuracy and system response adaptability of the intelligent instrument in complex dynamic measurement environments. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the execution flow of the intelligent instrument control method based on deep learning provided in an embodiment of the present invention.

[0008] Figure 2 This is a logical schematic diagram of the intelligent instrument control method based on deep learning provided in an embodiment of the present invention.

[0009] Figure 3 This is a schematic diagram of exemplary hardware and software components of a deep learning-based intelligent instrument control system provided in an embodiment of the present invention. Detailed Implementation

[0010] Figure 1 This is a flowchart illustrating a deep learning-based intelligent instrument control method according to an embodiment of the present invention. The following is a detailed explanation. Figure 2 A detailed introduction will be provided.

[0011] Step S110: Obtain the raw measurement signal data collected by the smart instrument in the current measurement cycle and the historical measurement signal data corresponding to the previous control cycle stored by the smart instrument.

[0012] In this embodiment, for an intelligent pressure transmitter in an industrial process control scenario, the original measurement signal data A collected by the intelligent instrument within the current measurement cycle t at a sampling frequency Fs is acquired. This original measurement signal data A is a one-dimensional time-series array of length L. Simultaneously, historical measurement signal data B corresponding to the previous control cycle t-1 is read from the intelligent instrument's internal memory. This historical measurement signal data B has the same data length L and sampling frequency Fs as A. All data acquisition processes have been authorized for on-site operation and do not involve any personal privacy information.

[0013] Step S120: Perform time-frequency domain joint decomposition processing on the original measurement signal data to generate measurement signal decomposition features containing time-domain component sequences and frequency-domain component sequences, and perform time-frequency domain joint decomposition processing on the historical measurement signal data to generate historical measurement signal decomposition features, wherein the historical measurement signal decomposition features contain historical time-domain component sequences and historical frequency-domain component sequences.

[0014] Step S121: Perform variable-scale time-domain window segmentation on the original measurement signal data to obtain a sequence of signal segments with temporal adjacency. Perform local oscillation mode extraction on each signal segment in the signal segment sequence to generate a set of narrowband oscillation components corresponding to each signal segment. The set of narrowband oscillation components consists of eigenmode components with a single dominant frequency oscillation attribute.

[0015] In this embodiment, variable-scale time-domain window segmentation is performed on the original measurement signal data A. A sliding window mechanism is adopted, with the window width W gradually increasing from a minimum to a maximum value, and the step size being a fixed value. The variance of the signal energy distribution is calculated under each window width, and the window width with the smallest variance is selected as the optimal window width Wopt. With Wopt as the fixed window width and the step size S equal to Wopt divided by 2, the original measurement signal data A is overlapped and segmented to generate P signal segments, denoted as U1, U2, ..., UP, where adjacent signal segments have a temporal adjacency relationship, and the overlap is half of Wopt. Then, local oscillation mode extraction is performed on each signal segment Ui. An empirical mode decomposition algorithm is used to decompose the signal segment Ui into multiple eigenmode components through a screening process. The selection process is as follows: Identify all local maxima and local minima of signal segment Ui; fit the upper and lower envelopes using cubic spline interpolation; calculate the mean envelope; subtract the mean envelope from signal segment Ui to obtain candidate components; repeat the above process until the difference between the number of extreme points and the number of zero-crossing points of the candidate components is less than or equal to 1, and the mean envelope of the candidate components is approximately zero, thus obtaining the first intrinsic mode component. Subtract this intrinsic mode component from signal segment Ui and repeat the above process to obtain subsequent intrinsic mode components. The set of intrinsic mode components generated for each signal segment Ui is denoted as Vi, and each intrinsic mode component in the set Vi has a single dominant frequency oscillation attribute.

[0016] Step S122: For each signal segment, select the intrinsic mode component with the largest change in zero crossover rate from the narrowband oscillation component set as the time-domain dominant component of the signal segment, and splice the time-domain dominant components of all signal segments in chronological order to generate the time-domain component sequence.

[0017] In this embodiment, for each intrinsic mode component set Vi corresponding to a signal segment Ui, the change in zero-crossing rate of each intrinsic mode component is calculated. The change in zero-crossing rate is calculated by counting the number of times the intrinsic mode component crosses a zero point per unit time, and summing the absolute values ​​of the zero-crossing rate differences within adjacent time windows to obtain the change value. The intrinsic mode component with the largest change value is selected as the dominant temporal component Wi of the signal segment Ui. The dominant temporal components W1, W2, ..., WP of all P signal segments are concatenated according to the temporal order of the signal segments to generate the temporal component sequence Wseq.

[0018] Step S123: Perform Hilbert spectrum mapping processing on each intrinsic mode component in the narrowband oscillation component set to obtain the instantaneous frequency change curve and instantaneous energy decay curve of each intrinsic mode component.

[0019] In this embodiment, for each eigenmode component in the eigenmode component set Vi corresponding to each signal segment Ui, Hilbert spectrum mapping processing is performed. A Hilbert transform is performed on the eigenmode component, i.e., the derivative of the phase function of the analytic signal formed by the eigenmode component and its Hilbert transform is calculated to obtain the instantaneous frequency change curve. Simultaneously, the square of the amplitude of the analytic signal is calculated to obtain the instantaneous energy decay curve. For each eigenmode component in the eigenmode component set Vi, a corresponding instantaneous frequency change curve and instantaneous energy decay curve are generated.

[0020] Step S124: Align the instantaneous frequency change curves of all intrinsic mode components within each signal segment along the time axis to generate the instantaneous frequency spectrum of that signal segment, and stitch together the instantaneous frequency spectra of all signal segments in chronological order to generate the frequency domain component sequence.

[0021] In this embodiment, for each signal segment Ui, the instantaneous frequency change curves of all intrinsic mode components in its intrinsic mode component set Vi are aligned along the same time axis to form a two-dimensional matrix. The rows of this two-dimensional matrix correspond to different intrinsic mode components, the columns correspond to time points, and the matrix elements are instantaneous frequency values. This two-dimensional matrix is ​​used as the instantaneous frequency spectrum Fi of the signal segment Ui. The instantaneous frequency spectra F1, F2, ..., FP of all P signal segments are spliced ​​together according to the time order of the signal segments to generate the frequency domain component sequence Fseq.

[0022] Step S125: Perform recursive graph generation processing on the time-domain component sequence, calculate the distance relationship between any two time points in the time-domain component sequence based on phase space reconstruction, and generate a time-domain recursive topology graph expressed in the form of a recursive matrix.

[0023] In this embodiment, the time-domain component sequence Wseq generated in step S122 is subjected to recursive graph generation processing. Using the phase space reconstruction method, the embedding dimension m and time delay τ are selected to reconstruct the time-domain component sequence Wseq into a set of points in the phase space, where the coordinates of each point are (Wseq(i), Wseq(i+τ), ..., Wseq(i+(m-1)*τ)). The distance between any two points after reconstruction is calculated, and the distance between point i and point j is denoted as Dij. A recursion threshold R is set; if Dij is less than R, the element in the i-th row and j-th column of the recursion matrix is ​​recorded as 1; otherwise, it is recorded as 0. The resulting two-dimensional recursion matrix Rmat serves as the time-domain recursive topology graph, reflecting the repetition pattern and periodic characteristics in the time-domain component sequence.

[0024] Step S126: Perform frequency band energy entropy extraction processing on the frequency domain component sequence, accumulate the instantaneous energy decay curve of each intrinsic mode component in the frequency domain component sequence by frequency point to generate a frequency point energy distribution curve, and calculate the proportion coefficient of each frequency point energy to the total energy based on the frequency point energy distribution curve, and generate a frequency band energy entropy vector by frequency point sequence.

[0025] In this embodiment, the frequency domain component sequence Fseq generated in step S124 is subjected to frequency band energy entropy extraction processing. For the instantaneous frequency spectrum Fi of each signal segment Ui, the instantaneous energy attenuation curves of each intrinsic mode component are accumulated by frequency point to obtain the frequency point energy distribution curve Ei(f) of the signal segment, where f represents the frequency. The total energy Etotal = ∑_fEi(f) of the frequency point energy distribution curve Ei(f) is calculated. For each frequency point f, the proportion coefficient Pf of its energy to the total energy is calculated as Pf = Ei(f) / Etotal. Then, the frequency band energy entropy value H = -∑_fPf*log(Pf) of the signal segment is calculated. The frequency band energy entropy values ​​of all signal segments are arranged in the order of the signal segments to generate the frequency band energy entropy vector Hvec.

[0026] Step S127: The time-domain recursive topology diagram and the frequency band energy entropy vector are respectively used as the time-domain component and frequency-domain component of the measurement signal decomposition feature. The measurement signal decomposition feature contains independent time-domain components and frequency-domain components, wherein the time-domain component in the measurement signal decomposition feature corresponds to the time-domain recursive topology diagram, and the frequency-domain component in the measurement signal decomposition feature corresponds to the frequency band energy entropy vector.

[0027] In this embodiment, the time-domain recursive topology diagram Rmat generated in step S125 is used as the time-domain component of the measurement signal decomposition feature, and the frequency band energy entropy vector Hvec generated in step S126 is used as the frequency-domain component of the measurement signal decomposition feature. The measurement signal decomposition feature Feat contains two independent components: the time-domain component Feat_t=Rmat and the frequency-domain component Feat_f=Hvec.

[0028] Step S128: Using the same processing flow, the historical measurement signal data is sequentially processed by variable-scale time-domain window segmentation, local oscillation mode extraction, time-domain dominant component selection, Hilbert spectrum mapping, instantaneous frequency spectrum generation, recursive graph generation, and frequency band energy entropy extraction to generate historical measurement signal decomposition features. The historical measurement signal decomposition features include independent historical time-domain components and historical frequency-domain components, wherein the historical time-domain components correspond to the historical time-domain recursive topology graph, and the historical frequency-domain components correspond to the historical frequency band energy entropy vector.

[0029] In this embodiment, the same processing flow as steps S121 to S127 is used to process the historical measurement signal data B. First, variable-scale time-domain windowing is performed on B to generate a sequence of signal segments. Empirical mode decomposition is performed on each signal segment to generate a set of eigenmode components. The eigenmode component with the largest change in zero crossover rate is selected as the dominant time-domain component and spliced ​​to generate a historical time-domain component sequence. Recursive graph generation processing is performed on this historical time-domain component sequence to obtain a historical time-domain recursive topology graph Rmat_hist. Simultaneously, Hilbert spectrum mapping is performed on the eigenmode components of each signal segment to obtain instantaneous frequency change curves and instantaneous energy attenuation curves, generating instantaneous frequency spectra and splicing them to generate a historical frequency-domain component sequence. Frequency band energy entropy extraction processing is performed on this historical frequency-domain component sequence to obtain a historical frequency band energy entropy vector Hvec_hist. Rmat_hist is used as the time-domain component Feat_hist_t of the historical measurement signal decomposition feature Feat_hist, and Hvec_hist is used as the frequency-domain component Feat_hist_f of the historical measurement signal decomposition feature Feat_hist.

[0030] Step S130: Input the measured signal decomposition features and the historical measured signal decomposition features into a pre-constructed deep temporal correlation enhancement network. Perform joint inference operations through the cross-period dependency extraction subnet and signal component interaction subnet within the deep temporal correlation enhancement network to generate a control parameter adjustment scheme.

[0031] Step S131: In the deep temporal correlation reinforcement network, the temporal recursive topology graph contained in the measurement signal decomposition features is input to the graph isomorphic inference layer of the cross-period dependency extraction subnet. The neighborhood structure of each node in the temporal recursive topology graph is traversed through the message passing mechanism to generate a graph isomorphic representation vector. The historical temporal recursive topology graph contained in the historical measurement signal decomposition features is input to the graph isomorphic inference layer of the cross-period dependency extraction subnet. The neighborhood structure of each node in the historical temporal recursive topology graph is traversed through the message passing mechanism to generate a historical graph isomorphic representation vector.

[0032] In this embodiment, a deep temporal correlation enhancement network is constructed, which includes a cross-period dependency extraction subnet and a signal component interaction subnet. The cross-period dependency extraction subnet includes a graph isomorphic inference layer and a cross-period attention aggregation layer. The graph isomorphic inference layer adopts a graph isomorphic network architecture, which includes multiple graph convolutional layers. The temporal recursive topology graph Rmat of the measured signal decomposition feature (Feat) is used as the graph structure input, where each node corresponds to a time point in the temporal component sequence, and the edges are determined by the non-zero elements in the recursion matrix. In the graph isomorphic inference layer, the neighborhood features of each node are aggregated through a message passing mechanism. The node feature update rule of the l-th layer is: h_v^{(l+1)}=MLP^{(l)}((1+ε^{(l)})*h_v^{(l)}+∑_{uinN(v)}h_u^{(l)}), where N(v) represents the set of neighborhood nodes of node v, ε^{(l)} is a learnable parameter, and MLP is a multilayer perceptron. After multi-layer graph convolution, global summation pooling is performed on the features of all nodes to obtain the graph isomorphic representation vector Z. The same processing is performed on the historical temporal recursive topological structure graph Rmat_hist to generate the historical graph isomorphic representation vector Z_hist.

[0033] Step S132: Input the graph isomorphic representation vector and the historical graph isomorphic representation vector into the cross-period attention aggregation layer of the cross-period dependency extraction subnet, calculate the periodic correlation matrix between the graph isomorphic representation vector and the historical graph isomorphic representation vector, and perform weighted aggregation on the historical graph isomorphic representation vector through the periodic correlation matrix to generate a cross-period temporal dependency vector.

[0034] In this embodiment, the graph isomorphic representation vector Z and Z_hist are input into the cross-period attention aggregation layer. First, the dot product between the transposes of Z and Z_hist is calculated to obtain the periodic correlation matrix M, where M(i,j) = Z_i * Z_hist_j. Softmax normalization is performed on each row of M to obtain the attention weight matrix A. Then, A is used to perform weighted aggregation on Z_hist to obtain the cross-period temporal dependency vector D = A * Z_hist. This cross-period temporal dependency vector D integrates the temporal dependencies between the current period and historical periods.

[0035] Step S133: Input the frequency band energy entropy vector contained in the measurement signal decomposition features into the frequency band relationship coding layer of the signal component interaction subnet, and perform relative position embedding coding on the frequency point positions in the frequency band energy entropy vector to generate a frequency band position coding vector.

[0036] In this embodiment, the band energy entropy vector Hvec of the measured signal decomposition feature (Feat) is input into the band relation coding layer of the signal component interaction subnet. This band relation coding layer performs relative position embedding coding on each frequency point position p in Hvec. The coding method is as follows: for the p-th frequency point, the 2k-th component of its position embedding vector PE(p) is equal to sin(p / 10000^(2k / d)), and the 2k+1-th component is equal to cos(p / 10000^(2k / d)), where d is the embedding dimension. The position embedding vectors of all frequency points are stacked to generate the band position coding vector Pvec.

[0037] Step S134: Input the frequency band energy entropy vector and the frequency band position coding vector into the frequency band self-attention layer of the signal component interaction subnet, calculate the frequency point association weights between each frequency point, and perform a weighted summation of the frequency band energy entropy vector according to the frequency point association weights to generate a frequency band context aggregation vector.

[0038] In this embodiment, the frequency band energy entropy vector Hvec and the frequency band location encoding vector Pvec are added element-wise to obtain the frequency band representation vector Q with fused location information. This vector is then input into the frequency band self-attention layer. The frequency band self-attention layer first maps Q to a query matrix Qq, a key matrix Qk, and a value matrix Qv through three different linear transformation matrices. The attention weight matrix Att = softmax((Qq*Qk^T) / √d) is calculated, where d is the feature dimension. Then, the frequency band context aggregation vector C = Att*Qv is calculated. This frequency band context aggregation vector C aggregates the context information between different frequency points.

[0039] Step S135: Input the historical frequency band energy entropy vector and the frequency band position encoding vector contained in the historical measurement signal decomposition features into the frequency band self-attention layer of the signal component interaction subnet, calculate the historical frequency point association weights between historical frequency points, and perform weighted summation on the historical frequency band energy entropy vector according to the historical frequency point association weights to generate a historical frequency band context aggregation vector.

[0040] In this embodiment, the historical frequency band energy entropy vector Hvec_hist of the historical measurement signal decomposition feature Feat_hist is element-wise added to the frequency band location encoding vector Pvec generated in step S133 to obtain the historical frequency band representation vector Q_hist of the fused location information. This vector is then input into the frequency band self-attention layer, and the same calculation process as in step S134 is used to generate the historical frequency band context aggregation vector C_hist through linear transformation, attention weight calculation, and weighted summation.

[0041] Step S136: Input the frequency band context aggregation vector and the historical frequency band context aggregation vector into the cross-cycle frequency band fusion layer of the signal component interaction subnet, calculate the interaction change of the frequency band context aggregation vector and the historical frequency band context aggregation vector at the same frequency point, and generate a frequency band evolution difference vector.

[0042] In this embodiment, the frequency band context aggregation vector C and the historical frequency band context aggregation vector C_hist are input into the cross-cycle frequency band fusion layer. The difference between the two vectors at the same frequency point is calculated to obtain the frequency band evolution difference vector D_freq = C - C_hist. This frequency band evolution difference vector D_freq reflects the frequency domain variation trend of the current cycle relative to the historical cycle.

[0043] Step S137: Input the cross-period temporal dependency vector and the frequency band evolution difference vector into the joint inference layer of the deep temporal correlation enhancement network, fuse the cross-period temporal dependency vector and the frequency band evolution difference vector into a joint inference input vector, and perform layer-by-layer dimensional transformation processing on the joint inference input vector through the fully connected computing unit of the joint inference layer to generate a control parameter adjustment vector, wherein each element in the control parameter adjustment vector corresponds to the numerical adjustment amount of the control parameter.

[0044] In this embodiment, the cross-cycle temporal dependency vector D generated in step S132 and the frequency band evolution difference vector D_freq generated in step S136 are input into the joint inference layer. First, D and D_freq are concatenated along the feature dimension to obtain the joint inference input vector U = concat(D, D_freq). The joint inference layer contains multiple fully connected layers, each containing a linear transformation and a nonlinear activation function. The first fully connected layer maps U to a vector of dimension h1, the second layer maps it to dimension h2, and the last layer maps it to a vector of dimension equal to the number of control parameters, denoted as ΔP. ​​Each element in this control parameter adjustment vector ΔP corresponds to the numerical adjustment amount of a control parameter.

[0045] Step S138: Add the corresponding elements of the control parameter adjustment vector to the current control parameters of the smart instrument in the current control cycle to generate an adjusted control parameter set, and use the adjusted control parameter set as the control parameter adjustment scheme.

[0046] In this embodiment, the current control parameter set Pcur within the current control cycle of the smart instrument is obtained, which includes the actuator control voltage value Vcur and signal acquisition module parameters Ccur (such as sampling frequency, gain, etc.). The control parameter adjustment vector ΔP generated in step S137 is added to Pcur element by element to obtain the adjusted control parameter set Pnew = Pcur + ΔP. Pnew is used as the control parameter adjustment scheme for the smart instrument.

[0047] Step S140: Generate an instrument control command set according to the control parameter adjustment scheme, and send the instrument control command set to the actuator and signal acquisition module through the control interface of the smart instrument. The instrument control command set includes actuator control voltage commands and signal acquisition module parameter commands.

[0048] In this embodiment, an instrument control command set is generated based on the control parameter adjustment scheme Pnew generated in step S138. This instrument control command set includes two commands: an actuator control voltage command C_act, which contains the adjusted actuator control voltage value; and a signal acquisition module parameter command C_sig, which contains the adjusted signal acquisition module parameters (such as sampling frequency, gain, etc.). Through the intelligent instrument's control interface, the actuator control voltage command C_act is sent to the actuator, and the signal acquisition module parameter command C_sig is sent to the signal acquisition module, causing the actuator and signal acquisition module to operate according to the adjusted parameters.

[0049] Step S210: Update the actuator control voltage and signal acquisition module parameters of the intelligent instrument according to the control parameter adjustment scheme, and obtain the corrected measurement signal data of the next measurement cycle under the drive of the updated control parameters. Perform time-frequency domain joint decomposition processing on the corrected measurement signal data to generate corrected measurement signal decomposition features. The corrected measurement signal decomposition features include a corrected time-domain recursive topology diagram and a corrected frequency band energy entropy vector.

[0050] In this embodiment, the actuator control voltage and signal acquisition module parameters of the intelligent instrument are updated according to the control parameter adjustment scheme Pnew generated in step S138. Driven by the updated control parameters, the intelligent instrument enters the next measurement cycle t+1 and acquires new corrected measurement signal data C. The time-frequency domain joint decomposition processing flow, which is exactly the same as steps S121 to S127, is performed on the corrected measurement signal data C to generate the corrected measurement signal decomposition feature Fnew. The corrected measurement signal decomposition feature Fnew includes the corrected time-domain recursive topology diagram Rnew and the corrected frequency band energy entropy vector Hnew.

[0051] Step S220: Input the frequency band energy entropy vector in the measured signal decomposition features and the modified frequency band energy entropy vector in the modified measured signal decomposition features to the frequency band response change evaluation module, and generate the frequency band response change vector by calculating the energy entropy change at the same frequency point.

[0052] In this embodiment, the frequency band energy entropy vector H generated in step S126 and the corrected frequency band energy entropy vector Hnew generated in step S210 are input to the frequency band response change evaluation module, which calculates the energy entropy change at the same frequency point using the formula D(i) = Hnew(i) - H(i), where i represents the frequency point index. The changes at all frequency points are then combined to form the frequency band response change vector Dvec.

[0053] Step S230: Perform correlation analysis on the energy entropy change of each control parameter in the adjusted control parameter set included in the control parameter adjustment scheme and the corresponding frequency point in the frequency band response change vector to generate a parameter frequency band response correlation degree sequence. Based on the sorting result of each correlation degree value in the parameter frequency band response correlation degree sequence, select a predetermined number of control parameters with the highest correlation degree as the target sensitive parameter set.

[0054] In this embodiment, the control parameter adjustment scheme Pnew contains M control parameters, denoted as Q1, Q2, ..., QM. The frequency band response change vector Dvec contains the changes at N frequency points. For each control parameter Qj, the Pearson correlation coefficient between it and the energy entropy change at each frequency point is calculated, resulting in a correlation vector Rj of length N. The element with the largest absolute value in Rj is taken as the frequency band response correlation value Rjmax of that control parameter. The Rjmax values ​​of all control parameters are sorted in descending order, and the top T control parameters are selected to form the target sensitive parameter set S.

[0055] Step S240: Perform linear mapping processing on the control parameters in the target sensitive parameter set and the frequency band response change vector to generate a parameter response mapping relationship model. Add a parameter response correction branch in the joint inference layer of the deep temporal correlation enhancement network and load the target sensitive parameter set and the parameter response mapping relationship model into the parameter response correction branch.

[0056] In this embodiment, for each control parameter Qj in the target sensitive parameter set S, a linear regression analysis is performed between it and the frequency band response change vector Dvec to establish a mapping relationship: Dvec = Aj * Qj + Bj, where Aj and Bj are regression coefficients. All mapping relationships are combined into a parameter response mapping relationship model Map. A parameter response correction branch is added to the joint inference layer of the deep temporal correlation reinforcement network, and the target sensitive parameter set S and the parameter response mapping relationship model Map are loaded into this branch.

[0057] Step S250: The elements belonging to the target sensitive parameter set in the control parameter adjustment vector generated by the joint inference layer are subjected to secondary adjustment processing based on the mapping relationship through the parameter response correction branch to generate the corrected control parameter adjustment vector. The corrected control parameter adjustment vector replaces the original control parameter adjustment vector as the updated control parameter adjustment scheme, and the updated control parameter adjustment scheme drives the next control cycle of the smart instrument.

[0058] In this embodiment, after the control parameter adjustment vector U is generated in the joint inference layer, the parameter response correction branch extracts elements belonging to the target sensitive parameter set S, denoted as Us. For each adjustment amount in Us, a secondary adjustment is performed using the corresponding mapping relationship in the parameter response mapping model Map: Usnew = Us - L * Aj, where L is the correction coefficient. The adjusted Usnew is written back to the corresponding position in U to generate the corrected control parameter adjustment vector Ucorr. Ucorr replaces the original U as the updated control parameter adjustment scheme, driving the next control cycle of the intelligent instrument.

[0059] Step S310: Obtain the control parameter adjustment scheme records generated by the smart instrument in multiple consecutive control cycles to form a control parameter adjustment history sequence. Perform parameter decoupling processing on the control parameter adjustment scheme corresponding to each control cycle in the control parameter adjustment history sequence, and extract the actuator control voltage adjustment amount and signal acquisition module parameter adjustment amount corresponding to each control cycle.

[0060] In this embodiment, the control parameter adjustment scheme records generated by the smart instrument within P consecutive control cycles are acquired, forming a control parameter adjustment history sequence Seq. For the control parameter adjustment scheme corresponding to each control cycle in Seq, parameter decoupling processing is performed to extract the actuator control voltage adjustment amount E and the signal acquisition module parameter adjustment amount F.

[0061] Step S320: Arrange the actuator control voltage adjustment amount in the continuous control cycle in chronological order to generate the actuator adjustment time series, and arrange the signal acquisition module parameter adjustment amount in the continuous control cycle in chronological order to generate the acquisition module parameter adjustment time series.

[0062] In this embodiment, the actuator control voltage adjustment amounts E of all control cycles are arranged in chronological order to obtain the actuator adjustment time sequence Te=[E1, E2, ..., EP]. The signal acquisition module parameter adjustment amounts F of all control cycles are arranged in chronological order to obtain the acquisition module parameter adjustment time sequence Tf=[F1, F2, ..., FP].

[0063] Step S330: Perform trend decomposition processing on the actuator adjustment time series, and decompose the actuator adjustment time series into control trend component series and control fluctuation component series through a sliding window.

[0064] In this embodiment, a sliding window averaging method is used to perform trend decomposition on the actuator adjustment time series Te. A window width W is set, and the mean within each window is calculated as the trend component, with the formula Gi = (1 / W) * ∑_{j=iW / 2}^{i+W / 2}Te(j). The trend component is subtracted from the original sequence to obtain the fluctuation component, with the formula Hi = Te(i) - Gi. All windows are traversed to obtain the control trend component sequence Gseq and the control fluctuation component sequence Hseq.

[0065] Step S340: Perform differential autoregressive analysis on the control trend component sequence to generate an actuator control trend prediction model, and predict the actuator control voltage adjustment trend value for the next control cycle using the actuator control trend prediction model. Perform conditional probability distribution modeling on the control fluctuation component sequence to generate a control fluctuation distribution model, and predict the actuator control voltage adjustment fluctuation range for the next control cycle using the control fluctuation distribution model.

[0066] In this embodiment, an ARIMA model is fitted to the control trend component sequence Gseq, and the model order is determined using the autocorrelation function and partial autocorrelation function to generate the actuator control trend prediction model Md1. Md1 is used to predict the trend value Vtr for the next control period. Kernel density estimation is performed on the control fluctuation component sequence Hseq to generate the control fluctuation distribution model Md2. Quantiles are extracted from Md2 to obtain the fluctuation prediction interval [Vlo, Vhi] for the next control period.

[0067] Step S350: The actuator control voltage adjustment trend value and the actuator control voltage adjustment fluctuation range are synthesized to generate an actuator control voltage adjustment prediction interval. When generating the control parameter adjustment scheme for the next control cycle, the actuator control voltage adjustment prediction interval is used as a constraint condition and input to the joint inference layer of the deep temporal correlation reinforcement network. The joint inference layer searches for a control parameter adjustment vector that meets the optimization index within the actuator control voltage adjustment prediction interval.

[0068] In this embodiment, the predicted trend value Vtr is synthesized with the fluctuation range [Vlo, Vhi] to obtain the actuator control voltage adjustment prediction range [Vtr+Vlo, Vtr+Vhi]. When generating the control parameter adjustment vector for the next control cycle in the joint inference layer of the deep temporal correlation reinforcement network, this prediction range is used as a constraint, requiring that the actuator control voltage adjustment amount in the generated control parameter adjustment vector must fall within this range.

[0069] Step S360: By introducing the actuator control voltage adjustment prediction interval as a constraint, the search space of the control parameters of the deep temporal correlation reinforcement network in the next control cycle is narrowed.

[0070] In this embodiment, in the joint inference layer of the deep temporal correlation reinforcement network, the actuator control voltage adjustment prediction interval [Vtr+Vlo, Vtr+Vhi] is encoded as an activation range constraint for the corresponding neurons in the output layer. When calculating the control parameter adjustment vector in the output layer, for the neuron corresponding to the actuator control voltage adjustment, its output value, after sigmoid activation, is linearly mapped to this prediction interval, rather than the original 0 to 1 range. This constraint significantly reduces the network's search range in the output space, accelerates convergence, and improves control stability.

[0071] Step S410: Obtain environmental sensing data of the target site environment where the smart instrument is located. The environmental sensing data includes temperature distribution data and vibration spectrum data of the target site.

[0072] In this embodiment, a temperature sensor array and a vibration sensor are deployed at the target industrial site where the smart instrument is located. The temperature sensor array collects temperature distribution data (Tmap) from the site, which is stored in the form of a grid matrix, with each grid point corresponding to a temperature value. The vibration sensor collects vibration spectrum data (Sfreq) from the site, which is stored in the form of a frequency-amplitude sequence with a length of K.

[0073] Step S420: Perform spatial temperature field interpolation processing on the temperature distribution data to generate a temperature field distribution map expressed in the form of a spatial grid, wherein each grid node of the temperature field distribution map corresponds to a temperature value.

[0074] In this embodiment, kriging interpolation is performed on the temperature distribution data Tmap to generate a temperature field distribution map Gtemp with higher spatial resolution. The size of the temperature field distribution map Gtemp is U times V, and each grid node corresponds to an interpolated temperature value.

[0075] Step S430: Perform dominant frequency component extraction processing on the vibration spectrum data. Extract a predetermined number of vibration dominant frequency components with the highest amplitude from the vibration spectrum data through amplitude spectrum analysis to generate a set of vibration dominant frequency components. Perform temperature gradient calculation processing on the temperature field distribution map. Generate a temperature gradient vector field by calculating the temperature difference between adjacent spatial grid nodes. Each vector in the temperature gradient vector field represents the direction and rate of temperature change at the corresponding location.

[0076] In this embodiment, amplitude spectrum analysis is performed on the vibration spectrum data Sfreq to extract the R frequency components with the largest amplitudes, forming the vibration dominant frequency component set Pvib. Each dominant frequency component contains a frequency value and a corresponding amplitude. Temperature gradient calculation is performed on the temperature field distribution map Gtemp. For each grid node (i, j), the temperature difference between it and its neighboring nodes (i+1, j) and (i, j+1) is calculated, resulting in the gradient vector grad(i, j) = (Gtemp(i+1, j) - Gtemp(i, j), Gtemp(i, j+1) - Gtemp(i, j)). The gradient vectors of all nodes constitute the temperature gradient vector field Vgrad.

[0077] Step S440: Input the temperature gradient vector field and the set of vibration dominant frequency components into the environmental state embedding coding network, and map the temperature gradient vector field and the set of vibration dominant frequency components to the environmental embedding vector space of the same dimension through a fully connected embedding layer to generate an environmental state embedding vector. Input the environmental state embedding vector into the environment adaptation branch of the deep temporal correlation reinforcement network, and perform multi-layer nonlinear transformation processing on the environmental state embedding vector to generate an environment adaptation modulation vector.

[0078] In this embodiment, an environment state embedding coding network is constructed, which contains two parallel fully connected embedding layers. The temperature gradient vector field Vgrad is flattened into a one-dimensional vector and input into the first fully connected embedding layer, mapping it to an embedding vector Etemp of dimension D. The vibration dominant frequency component set Pvib is flattened and input into the second fully connected embedding layer, mapping it to an embedding vector Evib of the same dimension D. Etemp and Evib are concatenated along the feature dimension to obtain the environment state embedding vector Een. This vector is input into the environment adaptation branch of the deep temporal correlation reinforcement network. This environment adaptation branch contains two fully connected layers. The first layer outputs a dimension of 2D, and the second layer outputs a dimension of D. ReLU activation is used in between, and the final output is the environment adaptation modulation vector Menv.

[0079] Step S450: In the joint inference layer of the deep temporal correlation enhancement network, the environment adaptation modulation vector, the cross-cycle temporal dependency vector, and the frequency band evolution difference vector are subjected to gated fusion processing. The corresponding elements of the cross-cycle temporal dependency vector and the frequency band evolution difference vector are weighted and modulated by each element in the environment adaptation modulation vector. The cross-cycle temporal dependency vector and the frequency band evolution difference vector after modulation by the environment adaptation modulation vector are input to the fully connected computing unit of the joint inference layer to generate a control parameter adjustment vector that integrates environmental factors.

[0080] In this embodiment, a gated fusion unit is set in the joint inference layer of the deep temporal correlation enhancement network. The environment adaptation modulation vector Menv is split into two gated weight vectors G1 and G2, corresponding to the cross-cycle temporal dependency vector D and the frequency band evolution difference vector Dfreq, respectively. The modulated cross-cycle temporal dependency vector Dmod = G1⊙D and the modulated frequency band evolution difference vector Dfreq_mod = G2⊙Dfreq are calculated, where ⊙ represents element-wise multiplication. Dmod and Dfreq_mod are concatenated and input into the fully connected computation unit of the joint inference layer. This fully connected computation unit contains three fully connected network layers and finally outputs the control parameter adjustment vector Uenv that fuses environmental factors.

[0081] Step S510: Obtain the control effect feedback data generated by the smart instrument after executing the instrument control command set. The control effect feedback data includes the time series of the deviation between the actual measured value of the controlled variable and the target set value.

[0082] In this embodiment, after the smart instrument executes the set of instrument control commands, it collects the actual measurement sequence Yreal of the controlled variable (such as pressure value) and obtains the corresponding target setpoint sequence Ytarget. The deviation time series Edev = Yreal - Ytarget is calculated, and the length of this deviation time series is the same as the number of sampling points in the measurement period.

[0083] Step S520: Perform deviation pattern encoding processing on the deviation time series, divide the deviation time series into positive deviation subsequences and negative deviation subsequences according to the deviation direction, and perform deviation amplitude distribution statistical processing on the positive deviation subsequences and the negative deviation subsequences respectively to generate positive deviation distribution histograms and negative deviation distribution histograms. Perform distribution morphology parameter extraction processing on the positive deviation distribution histograms and the negative deviation distribution histograms respectively to generate a deviation distribution morphology parameter set, the deviation distribution morphology parameter set including the skewness parameter and the kurtosis parameter of the deviation distribution.

[0084] In this embodiment, the deviation time series Edev is traversed, and deviation values ​​greater than 0 are assigned to the positive deviation subsequence Epos, while deviation values ​​less than 0 are assigned to the negative deviation subsequence Eneg. Histogram statistics are performed on Epos and Eneg respectively, and a fixed number of statistical intervals B are set to obtain the positive deviation distribution histogram Hpos and the negative deviation distribution histogram Hneg. The skewness parameter Skew=(1 / N)*∑((xi-μ) / σ)^3 and the kurtosis parameter Kurt=(1 / N)*∑((xi-μ) / σ)^4-3 are calculated for Hpos and Hneg respectively, where N is the number of samples, μ is the mean, and σ is the standard deviation. Skew_pos, Kurt_pos, Skew_neg, and Kurt_neg are combined into the deviation distribution morphological parameter set Pdist.

[0085] Step S530: Input the deviation time series into the deviation feature extraction layer of the feedback strategy generation network, perform temporal convolution processing on the deviation time series, and generate deviation temporal feature vector.

[0086] In this embodiment, a feedback policy generation network is constructed, which includes a bias feature extraction layer, a policy mapping layer, and a policy parameter generation layer. The bias feature extraction layer consists of three one-dimensional convolutional layers, with kernel sizes of 3, 5, and 3, and channel numbers of 16, 32, and 64, respectively. The bias time series Edev is input into this layer, and after convolution, batch normalization, and ReLU activation, global average pooling is used to obtain the bias time series feature vector Fdev, with a dimension of 64.

[0087] Step S540: Input the deviation distribution morphological parameter set and the deviation temporal feature vector into the policy mapping layer of the feedback policy generation network, perform concatenation and fusion processing on the deviation distribution morphological parameter set and the deviation temporal feature vector to generate a fused deviation feature vector, and perform policy category prediction processing on the fused deviation feature vector through the policy classifier of the policy mapping layer to generate a feedback policy category identifier, which indicates the correction direction of the control parameter adjustment.

[0088] In this embodiment, the deviation distribution morphological parameter set Pdist is concatenated with the deviation temporal feature vector Fdev to obtain the fused deviation feature vector Ffusion, which has a dimension of 68. The policy mapping layer contains a fully connected classifier with an input dimension of 68 and an output dimension of 3 (corresponding to three policy categories: positive correction, negative correction, and no correction). The class probabilities are output through the softmax function, and the class with the highest probability is taken as the feedback policy category identifier Lfb.

[0089] Step S550: Through the policy parameter generation layer of the feedback policy generation network, a policy correction parameter vector is generated with the fusion deviation feature vector and the feedback policy category identifier as input. Each element in the policy correction parameter vector corresponds to the correction magnitude of the control parameter adjustment scheme.

[0090] In this embodiment, the policy parameter generation layer adopts a conditional generation network structure. The fusion bias feature vector Ffusion and the one-hot encoded vector of the feedback policy category identifier Lfb are concatenated and input into two fully connected layers. The first layer outputs a dimension of 128, and the second layer outputs a dimension equal to the number of control parameters M, resulting in a policy correction parameter vector Rcorr. Each element in this policy correction parameter vector corresponds to the correction magnitude of a control parameter, with values ​​ranging from -1 to 1.

[0091] Step S560: The strategy correction parameter vector and the control parameter adjustment scheme generated by the deep temporal correlation reinforcement network in the current control cycle are fused element by element to generate the control parameter adjustment scheme after feedback correction.

[0092] In this embodiment, the strategy correction parameter vector Rcorr is fused element-wise with the control parameter adjustment vector U generated in the current control cycle. The fusion formula is Ufb = U + γ * Rcorr, where γ is the feedback fusion coefficient, ranging from 0 to 1. Ufb is used as the control parameter adjustment scheme after feedback correction to update the control parameters of the smart instrument.

[0093] Step S610: Obtain the multi-sample measurement signal data set collected by the signal acquisition module of the smart instrument under different sampling configurations, perform time-frequency domain joint decomposition processing on each sample in the multi-sample measurement signal data set, and generate the sample time-domain recursive topology diagram and sample frequency band energy entropy vector corresponding to each sample. Each sample in the multi-sample measurement signal data set corresponds to a sampling configuration.

[0094] In this embodiment, the signal acquisition module of the smart instrument supports Q different sampling configurations (such as different sampling frequencies and different gain combinations). For each sampling configuration, R samples are continuously acquired to obtain a multi-sample measurement signal data set Coll. Steps S121 to S127 are performed on each sample in Coll to generate a sample time-domain recursive topology graph Gs and a sample frequency band energy entropy vector Hs corresponding to each sample.

[0095] Step S620: Input the time-domain recursive topology diagram of samples corresponding to multiple consecutive control cycles under the same sampling configuration into the topology stability analysis layer of the sampling effect evaluation network, calculate the graph structure similarity between the time-domain recursive topology diagrams of samples in adjacent cycles, generate the topology stability evaluation value, and input the frequency band energy entropy vector of samples corresponding to multiple consecutive control cycles under the same sampling configuration into the frequency band coverage analysis layer of the sampling effect evaluation network, calculate the energy distribution coverage ratio of the sample frequency band energy entropy vector on the whole frequency band, and generate the frequency band coverage evaluation value.

[0096] In this embodiment, a sampling effect evaluation network is constructed, which includes a topology stability analysis layer and a frequency band coverage analysis layer. For a continuous R sample time-domain recursive topology graph G1, G2, ..., GR under the same sampling configuration, the topology stability analysis layer calculates the graph edit distance between adjacent graphs and converts it into a similarity Sim = 1 / (1 + GED). The average of all adjacent similarities is used as the topology stability evaluation value Stopo. For a continuous R sample frequency band energy entropy vector H1, H2, ..., HR under the same sampling configuration, the frequency band coverage analysis layer marks the frequency points in each Hi whose energy entropy value is greater than the threshold Th as valid frequency points, and calculates the proportion of the union of all valid frequency points to the entire frequency band, which is used as the frequency band coverage evaluation value Scov.

[0097] Step S630: The topology stability evaluation value and the frequency band coverage evaluation value are weighted and summed to generate a comprehensive sampling effect score. The comprehensive sampling effect scores corresponding to all sampling configurations are sorted and compared. The sampling configuration with the best comprehensive sampling effect score is selected from all sampling configurations as the optimal sampling configuration parameter. The optimal sampling configuration parameter is written into the configuration register of the signal acquisition module of the smart instrument, so that the signal acquisition module performs signal acquisition according to the optimal sampling configuration parameter in subsequent measurement cycles.

[0098] In this embodiment, weights α and β are set, with α + β = 1. The comprehensive sampling performance score Sscore = α * Stopo + β * Scov. Sscore is calculated for all Q sampling configurations, and the configuration with the highest Sscore is selected as the optimal sampling configuration parameter Copt. Copt is written to the configuration register of the signal acquisition module through the control interface of the intelligent instrument, ensuring that subsequent measurement cycles use the optimal sampling configuration for signal acquisition.

[0099] Step S640: After each generation of the control parameter adjustment scheme, the optimal sampling configuration parameter is compared with the signal acquisition module parameter instruction in the control parameter adjustment scheme. If the signal acquisition module parameter instruction in the control parameter adjustment scheme is inconsistent with the optimal sampling configuration parameter, the optimal sampling configuration parameter is used to replace the signal acquisition module parameter instruction.

[0100] In this embodiment, after each deep temporal correlation reinforcement network generates a control parameter adjustment scheme U, the signal acquisition module parameter instruction part Ccmd is parsed from U. Ccmd is then compared item by item with the optimal sampling configuration parameter Copt. If there is a difference between Ccmd and Copt, Ccmd is replaced with Copt, and the replaced adjustment scheme Uopt is issued as the final control parameter adjustment scheme for execution. This ensures that the signal acquisition module always operates under the optimal sampling configuration, while preserving the network's optimization results for other parameters such as actuator control voltage.

[0101] Step S710: After the deep temporal correlation reinforcement network performs multiple rounds of control cycle inference, the control performance log data of the smart instrument is obtained. The control performance log data includes the deviation sequence between the actual value and the target value of the controlled variable in each control cycle and the adjustment step sequence of the control parameter adjustment scheme.

[0102] In this embodiment, after the deep temporal association reinforcement network continuously executes P rounds of control cycle inference, control performance log data Log is exported from the internal memory of the smart instrument. This control performance log data Log contains two parts: a deviation sequence Eseq, which records the deviation between the actual value and the target value of the controlled variable in each control cycle; and an adjustment step size sequence Aseq, which records the sum of the absolute values ​​of the adjustment steps of each control parameter in the control parameter adjustment scheme for each control cycle.

[0103] Step S720: Perform performance degradation trend extraction processing on the deviation sequence, calculate the slope of the residual change of the deviation sequence in the continuous period by sliding linear fitting, generate the performance degradation slope sequence, and perform step size distribution statistical processing on the adjustment step size sequence to calculate the distribution variance and distribution mean of the adjustment step size in the continuous period, and generate step size distribution statistical features.

[0104] In this embodiment, a sliding window linear fitting is used for the deviation sequence Eseq. A window width W_fit is set, and linear regression is performed on the data points within each window to calculate the slope k_i of the fitted line. The slopes k_i of all windows are arranged in order to obtain the performance degradation slope sequence Kseq. For the step-size adjustment sequence Aseq, its mean μ_A and variance σ_A^2 over continuous periods are calculated as the step-size distribution statistical feature Stat_A.

[0105] Step S730: Perform joint performance diagnosis processing on the performance degradation slope sequence and the step size distribution statistical features, and map the performance degradation slope sequence and the step size distribution statistical features to a performance degradation type identifier through a preset performance diagnosis mapping function.

[0106] In this embodiment, a performance diagnostic mapping function F_diag is constructed, with the input being the performance degradation slope sequence Kseq and the step size distribution statistical feature Stat_A. First, the mean μ_K and trend value (the difference ΔK between the slope of the last window and the slope of the first window) of Kseq are calculated. The diagnostic rules are as follows: if μ_K is greater than the first threshold and ΔK is positive, output identifier 1 indicates rapid degradation; if μ_K is less than the second threshold and ΔK is negative, output identifier 2 indicates slow improvement; if σ_A^2 is greater than the third threshold, output identifier 3 indicates oscillating degradation; otherwise, output identifier 0 indicates stable degradation. The output identifier is denoted as Tdeg.

[0107] Step S740: Select a corresponding network structure parameter adjustment strategy from the pre-built deep temporal correlation reinforcement network parameter library according to the performance degradation type identifier. The network structure parameter adjustment strategy includes an adjustment instruction for the number of graph isomorphic inference layer network layers of the cross-period dependency extraction subnet in the deep temporal correlation reinforcement network and an adjustment instruction for the number of attention heads of the frequency band self-attention layer of the signal component interaction subnet.

[0108] In this embodiment, a deep temporal correlation reinforcement network parameter library is pre-constructed, which includes a mapping table that maps the performance degradation type identifier Tdeg to network structure parameter adjustment strategies. If Tdeg is 1 (rapid degradation), the adjustment strategy is: increase the number of network layers in the graph isomorphic inference layer by 1, and increase the number of attention heads in the bandgap self-attention layer by 2. If Tdeg is 2 (slow improvement), the adjustment strategy is: keep the number of layers unchanged, and reduce the number of attention heads by 1. If Tdeg is 3 (oscillating degradation), the adjustment strategy is: increase the number of graph isomorphic inference layers by 2, and simultaneously increase the number of attention heads by 1. If Tdeg is 0 (stable), the adjustment strategy is: keep the existing structure unchanged. The adjustment instruction is denoted as Cadj.

[0109] Step S750: The adjustment instructions for the number of layers in the graph isomorphic inference layer and the adjustment instructions for the number of attention heads in the frequency band self-attention layer are written into the network configuration file of the deep temporal correlation reinforcement network. This triggers the deep temporal correlation reinforcement network to perform a hot update operation on the structural parameters. After the hot update operation is completed, the original measurement signal data is re-processed with time-frequency domain joint decomposition to generate a hot-updated time-domain recursive topology graph and a hot-updated frequency band energy entropy vector. These are then input into the hot-updated deep temporal correlation reinforcement network to generate a hot-updated control parameter adjustment scheme.

[0110] In this embodiment, the adjustment instruction Cadj is written into the network configuration file of the deep temporal correlation reinforcement network. A hot update operation is triggered: first, the forward inference of the network is paused, releasing the memory occupied by the original network structure. The graph isomorphic inference layer and the frequency band self-attention layer are reconstructed according to the adjustment instruction, the weight parameters of the original unmodified layers are loaded, and the newly added layers are randomly initialized. After the structure reconstruction is completed, the original measurement signal data X for the current period is reacquired, and steps S121 to S127 are executed to generate the hot-updated temporal recursive topology graph Rnew and the hot-updated frequency band energy entropy vector Hnew. Rnew and Hnew are input into the deep temporal correlation reinforcement network with hot-updated structural parameters. Following steps S131 to S138, a hot-updated control parameter adjustment scheme Unew is generated, and Unew replaces the original scheme to continue subsequent control.

[0111] For example, the method may further include: step S810: acquiring the communication data stream between the smart instrument and other associated smart instruments in the target industrial site, the communication data stream including the associated instrument measurement signal data and the associated instrument control status data sent by the associated smart instrument.

[0112] In this embodiment, multiple smart meters are interconnected via an industrial Ethernet network in the target industrial site. A data acquisition gateway acquires the communication data flow (Dataflow) between the current smart meter and other associated smart meters. This Dataflow includes measurement signal data (Z) sent by the associated meter and control status data (State) from other associated meters (such as current actuator voltage, sampling frequency, etc.).

[0113] Step S820: Perform the same time-frequency domain joint decomposition process on the associated instrument measurement signal data as on the original measurement signal data to generate a time-domain recursive topology diagram of the associated instrument and a frequency band energy entropy vector of the associated instrument.

[0114] In this embodiment, the time-frequency domain joint decomposition processing flow is exactly the same as steps S121 to S127 on the associated instrument measurement signal data Z, generating the associated instrument time-domain recursive topology diagram Rrel and the associated instrument frequency band energy entropy vector Hrel.

[0115] Step S830: Input the time-domain recursive topology diagram of the associated instrument into the collaborative control inference subnet added to the deep temporal correlation enhancement network. Perform graph matching processing on the time-domain recursive topology diagram of the associated instrument and the time-domain recursive topology diagram of the smart instrument through the cross-instrument graph isomorphism layer to generate a cross-instrument graph matching degree matrix. Through the graph matching aggregation layer, perform graph feature aggregation processing on the time-domain recursive topology diagram of the associated instrument with the cross-instrument graph matching degree matrix as weight to generate an instrument collaborative topology representation vector.

[0116] In this embodiment, a collaborative control inference subnetwork is added to the deep temporal correlation enhancement network. This subnetwork includes a cross-instrument graph isomorphism layer and a graph matching aggregation layer. The cross-instrument graph isomorphism layer adopts a graph matching network architecture, taking the temporal recursive topology graph Rrel of the associated instrument and the temporal recursive topology graph Rmat of the current smart instrument as input, and calculating the node matching degree matrix Wmatch between the two graphs, where Wmatch(i,j)=cosine_similarity(h_i,h_j), and h_i and h_j are the embedding vectors of nodes in the two graphs, respectively. The graph matching aggregation layer uses Wmatch as the weight to perform weighted aggregation of the graph node features of Rrel, obtaining the instrument collaborative topology representation vector Vcoop.

[0117] Step S840: Calculate the covariance between the frequency band energy entropy vector of the associated instrument and the frequency band energy entropy vector of the smart instrument at the same frequency point, and generate a frequency band cooperative response vector.

[0118] In this embodiment, the covariance at the same frequency point is calculated for the band energy entropy vector Hrel of the associated instrument and the band energy entropy vector H of the smart instrument. For each frequency point i, the covariance Cov_i = E[(Hrel(i) - μ_rel)(H(i) - μ)], where μ_rel and μ are the means of Hrel and H, respectively. The covariance values ​​of all frequency points are used to form the band cooperative response vector Vcov.

[0119] Step S850: Input the instrument cooperative topology representation vector and the frequency band cooperative response vector into the cooperative strategy generation layer of the cooperative control inference subnet, and input the concatenated instrument cooperative topology representation vector and the frequency band cooperative response vector into the fully connected inference unit to generate the cooperative control modulation vector.

[0120] In this embodiment, the cooperative strategy generation layer includes a fully connected inference unit. The instrument cooperative topology representation vector Vcoop and the band cooperative response vector Vcov are concatenated along the feature dimension to obtain a joint vector Vjoint = concat(Vcoop, Vcov). Vjoint is then input into the fully connected inference unit, which contains two fully connected layers. The first layer outputs a dimension of 256, and the second layer outputs a dimension equal to the number of control parameters M. The final output is a cooperative control modulation vector Mcoop, which is used to modulate subsequent control parameter adjustments.

[0121] Step S860: The cooperative control modulation vector is input to the joint inference layer of the deep temporal correlation reinforcement network. During the process of generating the control parameter adjustment vector in the joint inference layer, the joint inference input vector is element-wise modulated by the cooperative control modulation vector, so that the control parameter adjustment vector integrates the cooperative control information of the associated smart instruments in the target industrial site.

[0122] In this embodiment, in the joint inference layer of the deep temporal correlation reinforcement network, after obtaining the joint inference input vector U = concat(D, Dfreq), a cooperative control modulation vector Mcoop is introduced for element-wise modulation. The modulation method is: Umod = U⊙(1 + Mcoop), where ⊙ represents element-wise multiplication. The modulated joint inference input vector Umod incorporates the cooperative control information of the associated instruments. Then, Umod is input into the fully connected computing unit of the joint inference layer to generate the control parameter adjustment vector Ucoop according to step S137. This Ucoop is based not only on the measurement signal characteristics of the instruments themselves, but also on the operating status of the associated instruments in the target industrial site, realizing multi-instrument cooperative control.

[0123] Step S910: Obtain the set of paired samples of measurement signal decomposition features and control parameter adjustment schemes accumulated by the smart instrument in multiple historical running periods. Divide the set of paired samples into an early training sample subset and a recent training sample subset according to the time order of the running periods. Perform inverse reconstruction operation of time-frequency domain joint decomposition processing on the measurement signal decomposition features of each paired sample in the early training sample subset. Reconstruct the time-domain waveform of the original measurement signal data through the time-domain component sequence to generate the time-domain sequence of the early reconstructed signal.

[0124] In this embodiment, a set of paired samples (SetPair) of measurement signal decomposition features and control parameter adjustment schemes accumulated by the smart instrument over multiple historical operating periods is obtained. SetPair is divided into an early training sample subset (SubE) and a recent training sample subset (SubR) according to the time sequence of the operating periods. For each sample in SubE, the time-domain component sequence (Wseq) of its measurement signal decomposition features is extracted, and an inverse reconstruction operation of time-frequency domain joint decomposition processing is performed. The inverse reconstruction process is as follows: Wseq is used as a superposition of eigenmode components, and through phase reconstruction and amplitude modulation, the time-domain waveform shape of the original measurement signal data is reconstructed to generate the early reconstructed signal time-domain sequence (RecE).

[0125] Step S920: Perform an inverse reconstruction operation of time-frequency domain joint decomposition processing on the measurement signal decomposition features of each paired sample in the recent training sample subset, and reconstruct the time-domain waveform of the original measurement signal data through the time-domain component sequence to generate the time-domain sequence of the recently reconstructed signal.

[0126] In this embodiment, for each sample in the recent training sample subset SubR, the same inverse reconstruction operation as in step S910 is used to generate the recent reconstruction signal time domain sequence RecR.

[0127] Step S930: Perform waveform complexity analysis on the early reconstructed signal time-domain sequence and the recent reconstructed signal time-domain sequence respectively. By calculating the zero-crossing density and the coefficient of variation of the adjacent peak interval of the reconstructed signal time-domain sequence, generate the early waveform complexity index and the recent waveform complexity index.

[0128] In this embodiment, for the reconstructed signal time-domain sequence RecE, its zero-crossing density Dzc = (number of zero-crossings) / (sequence length) is calculated. Simultaneously, peak points in the sequence are detected, and the coefficient of variation CV between adjacent peaks is calculated as (standard deviation of peak interval) / (mean of peak interval). The zero-crossing density and coefficient of variation are combined to form the early waveform complexity index CpxE = [Dzc_E, CV_E]. The same method is used for the recently reconstructed signal time-domain sequence RecR to generate the recent waveform complexity index CpxR = [Dzc_R, CV_R].

[0129] Step S940: Input the early waveform complexity index and the recent waveform complexity index into the complexity difference quantization layer of the concept drift detection module, calculate the complexity difference between the early waveform complexity index and the recent waveform complexity index through the complexity difference quantization layer, and generate a concept drift warning sign when the complexity difference exceeds a preset drift discrimination threshold.

[0130] In this embodiment, a concept drift detection module is constructed, which includes a complexity difference quantization layer. The complexity difference degree Diff = ||CpxR - CpxE||_2, that is, the Euclidean distance between two complexity index vectors, is calculated. A drift discrimination threshold Th_drift is set. If Diff > Th_drift, a concept drift warning flag Warn_drift is generated, indicating that the statistical characteristics of the recent signal have changed significantly from the earlier signal, that is, concept drift exists.

[0131] Step S950: Trigger the incremental learning process of the deep temporal correlation reinforcement network according to the concept drift warning flag. In the incremental learning process, the recent training sample subset is used as the incremental learning sample. By keeping the weights of the first few layers of the cross-period dependency extraction subnet of the deep temporal correlation reinforcement network unchanged, the weights of the last few layers of the cross-period dependency extraction subnet and the weights of the signal component interaction subnet are fine-tuned and updated based on the incremental learning sample.

[0132] In this embodiment, the incremental learning process is triggered when Warn_drift is true. The recently trained sample subset SubR is used as the incremental learning sample. During incremental learning, the weights of the first L layers of the cross-period dependency extraction subnet in the deep temporal correlation reinforcement network (L is half the total number of layers) are frozen and kept unchanged. The weights of the last few layers of the cross-period dependency extraction subnet and the weights of the entire signal component interaction subnet are fine-tuned using the incremental learning sample. A small learning rate η_inc (one-tenth of the original learning rate) is used, the optimizer is Adam, the training epochs are E_inc, and gradient descent updates are performed on SubR.

[0133] Step S960: Write the fine-tuned and updated cross-cycle dependency extraction subnet weights and the fine-tuned and updated signal component interaction subnet weights back to the network parameter file of the deep temporal correlation reinforcement network, replacing the original corresponding network weight parameters. After completing the network weight parameter replacement, perform time-frequency domain joint decomposition processing on the original measurement signal data of the next control cycle to generate incrementally updated measurement signal decomposition features. Input the incrementally updated measurement signal decomposition features into the deep temporal correlation reinforcement network after network weight replacement to generate a control parameter adjustment scheme adapted to concept drift.

[0134] In this embodiment, the updated network weights obtained through incremental learning are written back to the network parameter file of the deep temporal correlation reinforcement network, replacing the original corresponding weight parameters, thus completing the network hot update. In the next control cycle, new original measurement signal data Xnew is acquired, and steps S121 to S127 are executed to generate the incrementally updated measurement signal decomposition feature Feat_inc. Feat_inc is input into the deep temporal correlation reinforcement network after the network weights are replaced, and steps S131 to S138 are followed to generate the control parameter adjustment scheme U_inc adapted to concept drift. This control parameter adjustment scheme can effectively address the problem of signal statistical characteristic drift caused by environmental changes or equipment aging in industrial processes.

[0135] Step S1010: Obtain the actuator response current waveform data and actuator displacement feedback waveform data generated by the actuator of the smart instrument after receiving the control parameter adjustment scheme within a continuous control cycle.

[0136] In this embodiment, a current sensor and a displacement sensor are deployed at the actuator end of the smart instrument. The actuator response current waveform data Curr and displacement feedback waveform data Pos are collected after the actuator receives the control parameter adjustment scheme over Q consecutive control cycles. Both Curr and Pos are time-series data, and the sampling frequency is the same as the control cycle.

[0137] Step S1020: Perform pulse response segmentation processing on the actuator response current waveform data, and cut the actuator response current waveform data into multiple current response segments according to the time point of change of control command. Each current response segment corresponds to a current transient response process of one control parameter adjustment.

[0138] In this embodiment, the actuator response current waveform data Curr is processed using impulse response segmentation. First, the time point at which the control command change occurs is identified, i.e., the timestamp T_cmd when the control parameter adjustment scheme is issued. Using each T_cmd as the center, current response segments Frag_i are cut out by extending time windows W_resp forward and backward. Each current response segment Frag_i corresponds to one transient current response process during control parameter adjustment.

[0139] Step S1030: Perform damped oscillation feature extraction processing on each current response segment. By extracting the decay envelope curve and oscillation period sequence of the decaying oscillation process in the current response segment, the damped oscillation feature of the current response is generated.

[0140] In this embodiment, damped oscillation feature extraction is performed on each current response segment Frag_i. First, the oscillation waveform in Frag_i is detected, and peaks and troughs are identified. An exponential decay fit is performed on the peak sequence to obtain the decay envelope curve Enve(t) = A0*exp(-ξ*t), where ξ is the decay coefficient. The time intervals between adjacent peaks are extracted to form the oscillation period sequence Tosc. The decay coefficient ξ, the mean, and the variance of the oscillation period sequence Tosc are combined to form the current response damped oscillation feature Feat_i.

[0141] Step S1040: Perform the same pulse response segmentation processing on the actuator displacement feedback waveform data as on the actuator response current waveform data to generate displacement response segments, and perform steady-time extraction processing on each displacement response segment. By calculating the time length required for the displacement in the displacement response segment to reach the preset allowable deviation range of the target displacement value, a displacement response steady-time sequence is generated.

[0142] In this embodiment, the actuator displacement feedback waveform data Pos undergoes the same impulse response segmentation processing as in step S1020 to obtain displacement response segments Disp_i. For each displacement response segment Disp_i, its steady-state value (average value at the end of the segment) is extracted, and an allowable deviation range of ±δ is set. The time required for the displacement response to first enter and remain within the range of [steady-state value - δ, steady-state value + δ] from the starting point is calculated as the settling time Tstab_i of that segment. The settling times of all segments constitute the displacement response steady-state time sequence SeqTstab.

[0143] Step S1050: Input the current response damped oscillation characteristics and the displacement response stable time series into the time-series degradation modeling layer of the actuator health assessment network, perform time-series modeling on the changing trends of the decay envelope curve slope and oscillation period sequence in the current response damped oscillation characteristics to generate a current response degradation trend vector, and perform time-series modeling on the changing trends of the displacement response stable time series to generate a displacement response degradation trend vector.

[0144] In this embodiment, an actuator health assessment network is constructed, which includes a temporal degradation modeling layer and a health fusion inference layer. The temporal degradation modeling layer adopts a two-stream LSTM structure. The input of the first-stream LSTM is the decay coefficient ξ and the oscillation period Tosc sequence in the damped oscillation feature Feat_i of the current response for each control cycle, and the output is the current response degradation trend vector Vec_cur. The input of the second-stream LSTM is the displacement response stable time series SeqTstab, and the output is the displacement response degradation trend vector Vec_pos.

[0145] Step S1060: Input the current response degradation trend vector and the displacement response degradation trend vector into the health fusion inference layer of the actuator health assessment network, perform dual-stream cross-attention fusion processing on the current response degradation trend vector and the displacement response degradation trend vector to generate a comprehensive actuator health assessment vector, input the comprehensive actuator health assessment vector into the remaining lifetime prediction layer of the actuator health assessment network, and map the comprehensive actuator health assessment vector into a predicted value of the remaining effective working cycles of the actuator through the remaining lifetime prediction layer.

[0146] In this embodiment, the health fusion inference layer employs a cross-attention mechanism. Using Vec_cur as the query and Vec_pos as the key and value, cross-attention features are calculated; simultaneously, using Vec_pos as the query and Vec_cur as the key and value, reverse cross-attention features are calculated. These two attention features are concatenated with the original vector and fused through a fully connected layer to obtain the comprehensive health assessment vector Vhealth. The remaining lifespan prediction layer consists of two fully connected layers and one output layer. Vhealth is input to the first fully connected layer (output dimension 64), activated by ReLU, and then input to the second fully connected layer (output dimension 32). Finally, the output layer maps it to the scalar value RUL, representing the predicted number of remaining effective working cycles of the actuator.

[0147] Step S1070: Input the predicted value of the remaining effective working cycles of the actuator into the joint inference layer of the deep temporal correlation reinforcement network. During the process of generating the control parameter adjustment vector in the joint inference layer, use the predicted value of the remaining effective working cycles of the actuator as the attenuation constraint factor of the control parameter adjustment amplitude, so that the generated control parameter adjustment amplitude gradually shrinks as the remaining effective working cycles of the actuator decreases.

[0148] In this embodiment, in the joint inference layer of the deep temporal correlation reinforcement network, before generating the control parameter adjustment vector U, the predicted remaining effective working cycles of the actuator, RUL, are introduced as a decay constraint. First, RUL is normalized: the rated lifespan cycle number RUL_init of the actuator is obtained, and the normalized decay factor α_norm = RUL / RUL_init is calculated. α_norm is a dimensionless parameter, ranging from 0 to 1. At the output of the joint inference layer, the originally generated control parameter adjustment amount U_raw is multiplied by the normalized decay factor α_norm to obtain the final control parameter adjustment vector U_health = α_norm * U_raw. Since α_norm is dimensionless, the dimension of U_raw is the original dimension of the control parameter adjustment amount. The product does not change the dimension, ensuring dimensional consistency. This allows the adjustment amplitude of the control parameters to automatically decrease when the actuator approaches the end of its lifespan, avoiding excessive control commands to the aging actuator.

[0149] Step S1080: By introducing the predicted value of the remaining effective working cycles of the actuator as the attenuation constraint factor, the adjustment magnitude of each element in the control parameter adjustment vector is weighted and attenuated in the joint inference layer to generate a control parameter adjustment vector that integrates the actuator health status constraint.

[0150] In this embodiment, the attenuation constraint operation of step S1070 is performed in the joint inference layer of each control cycle. The generated control parameter adjustment vector U_health, which integrates the actuator health status constraint, satisfies both the control optimization requirements driven by the current measurement signal and takes into account the health status of the actuator itself, thus achieving adaptive control parameter adjustment. U_health is then sent to the actuator as the final control parameter adjustment scheme.

[0151] Step S1110: Obtain the multi-source heterogeneous sensing data stream of the target industrial site where the smart instrument is located. The multi-source heterogeneous sensing data stream includes the vibration sensing time-series data of the smart instrument and the fluid pressure pulsation time-series data of the pipeline where the smart instrument is located.

[0152] In this embodiment, vibration sensors and pressure pulsation sensors are deployed at the target industrial site where the smart instrument is located. The vibration sensor is mounted on the casing of the smart instrument and collects vibration sensing time-series data Vdata, which is a time-amplitude sequence. The pressure pulsation sensor is mounted on the process pipeline where the smart instrument is located and collects fluid pressure pulsation time-series data Pdata, which is also a time-amplitude sequence. The sampling frequencies of the two sequences are synchronized, and the sampling period is consistent with the control period.

[0153] Step S1120: Perform time-frequency domain joint decomposition processing on the vibration sensing time series data to generate a vibration time-domain recursive topology diagram and a vibration frequency band energy entropy vector, and perform time-frequency domain joint decomposition processing on the fluid pressure pulsation time series data to generate a pressure pulsation time-domain recursive topology diagram and a pressure pulsation frequency band energy entropy vector.

[0154] In this embodiment, the vibration sensing time-series data Vdata undergoes the same time-frequency domain joint decomposition processing procedure as steps S121 to S127, generating a vibration time-domain recursive topology diagram R_vib and a vibration frequency band energy entropy vector H_vib. The same processing procedure is performed on the fluid pressure pulsation time-series data Pdata, generating a pressure pulsation time-domain recursive topology diagram R_pre and a pressure pulsation frequency band energy entropy vector H_pre.

[0155] Step S1130: Input the vibration time-domain recursive topology diagram and the pressure pulsation time-domain recursive topology diagram into the directed graph construction layer of the multi-source causal inference network, map the vibration time-domain recursive topology diagram into vibration graph node feature vectors, and map the pressure pulsation time-domain recursive topology diagram into pressure graph node feature vectors; construct fully connected directed edges between the vibration graph node feature vectors and the pressure graph node feature vectors to generate a heterogeneous directed fully connected graph structure, input the heterogeneous directed fully connected graph structure into the causal direction discrimination layer of the multi-source causal inference network, perform conditional independence testing on each directed edge in the heterogeneous directed fully connected graph structure, remove directed edges that do not satisfy the conditional independence assumption, and generate a heterogeneous causal skeleton graph.

[0156] In this embodiment, a multi-source causal inference network is constructed, which includes a directed graph construction layer, a causal direction discrimination layer, and a causal intensity estimation layer. The directed graph construction layer first maps R_vib to a vibration graph node feature vector matrix F_vib through a graph convolutional network, and maps R_pre to a pressure graph node feature vector matrix F_pre. Then, directed edges (bidirectional) are constructed between each node in F_vib and each node in F_pre, forming a heterogeneous directed fully connected graph structure G_full. The causal direction discrimination layer performs a conditional independence test on each directed edge in G_full, using the partial correlation coefficient test to determine whether two variables are independent given other variables. If they are conditionally independent, the directed edge is deleted; otherwise, it is retained. The retained directed edges after the test constitute a heterogeneous causal skeleton graph G_causal.

[0157] Step S1140: Input the heterogeneous causal skeleton graph into the causal strength estimation layer of the multi-source causal inference network, perform causal strength calculation processing based on information entropy on the directed edges retained in the heterogeneous causal skeleton graph, and generate the causal strength coefficient corresponding to each retained directed edge.

[0158] In this embodiment, the causal strength estimation layer calculates the causal strength coefficient for each retained directed edge X→Y in the heterogeneous causal skeleton graph G_causal. The calculation formula is: Strength(X→Y)=MI(X,Y)-min_{Z}MI(X,Y|Z), where MI is the mutual information and Z is the set of parent nodes of X and Y. After calculating the causal strength coefficient for all retained directed edges, the weighted heterogeneous causal skeleton graph G_weight is obtained.

[0159] Step S1150: Input the vibration frequency band energy entropy vector and the pressure pulsation frequency band energy entropy vector into the frequency domain causal alignment layer of the multi-source causal inference network, calculate the transfer entropy between the vibration frequency band energy entropy vector and the pressure pulsation frequency band energy entropy vector at the same frequency point, and generate a frequency domain transfer entropy matrix.

[0160] In this embodiment, the frequency domain causal alignment layer calculates the transfer entropy TE_i=I(H_vib(i);H_pre(i)|H_vib(i-1)) at the same frequency point i for the vibration frequency band energy entropy vector H_vib and the pressure pulsation frequency band energy entropy vector H_pre, where I represents conditional mutual information. The transfer entropy values ​​of all frequency points are organized into a frequency domain transfer entropy matrix M_te, which reflects the information transfer direction of vibration and pressure pulsation at different frequency points.

[0161] Step S1160: Using the frequency domain transfer entropy matrix as a constraint, perform frequency domain weighted modulation processing on the causal intensity coefficients of the directed edges in the heterogeneous causal skeleton graph. Multiply the frequency point weights corresponding to the transfer entropy values ​​in the frequency domain transfer entropy matrix into the causal intensity coefficients corresponding to the directed edges to generate a frequency domain weighted causal intensity graph. Input the frequency domain weighted causal intensity graph into the causal perception fusion branch added to the deep temporal correlation reinforcement network. Perform graph convolution aggregation processing on the frequency domain weighted causal intensity graph through the causal perception fusion branch to generate a multi-source causal context vector.

[0162] In this embodiment, for each directed edge in the heterogeneous causal skeleton graph G_weight, its associated frequency index is obtained (determined based on the dominant frequency range of the signals corresponding to the two variables connected by the edge), and the transfer entropy value of that frequency point in the frequency domain transfer entropy matrix M_te is extracted as a weight factor w_freq. The causal strength coefficient is updated to Strength_new=w_freq*Strength_old, generating a frequency domain weighted causal strength graph G_freq. This graph is input into the causal-aware fusion branch added to the deep temporal correlation reinforcement network. This causal-aware fusion branch contains two layers of graph convolutional networks, which perform graph convolutional aggregation on G_freq to obtain the multi-source causal context vector V_causal.

[0163] Step S1170: In the joint inference layer of the deep temporal correlation reinforcement network, the multi-source causal context vector, the cross-period temporal dependency vector, and the frequency band evolution difference vector are fused by tensor product to generate a joint inference input vector that fuses multi-source causal constraints, and the control parameter adjustment vector that fuses multi-source causal information is generated through the fully connected computing unit of the joint inference layer.

[0164] In this embodiment, in the joint inference layer, the multi-source causal context vector V_causal is fused with the cross-cycle temporal dependency vector D and the frequency band evolution difference vector D_freq through tensor product. First, the fused intermediate tensor T_fuse = V_causal ⊗ D is calculated, where ⊗ represents the outer product operation, resulting in a matrix. Then, this matrix is ​​flattened and concatenated with D_freq to obtain the joint inference input vector U_causal that fuses multi-source causal constraints. U_causal is input into the fully connected computing unit (three-layer fully connected network) of the joint inference layer, and the control parameter adjustment vector U_final is generated according to step S137. This control parameter adjustment vector fuses the causal constraint information of multi-source heterogeneous sensor data such as vibration and pressure pulsation, and can more accurately adapt to complex industrial field environments.

[0165] Step S1210: Obtain the internal inference state record data of the deep temporal correlation reinforcement network when generating control parameter adjustment schemes under multiple different control scenarios. The internal inference state record data includes the graph node embedding vector sequence of the graph isomorphic inference layer of the cross-period dependency extraction subnet in each inference step and the attention weight matrix sequence of the frequency band self-attention layer of the signal component interaction subnet in each inference step.

[0166] In this embodiment, an internal state recording function is enabled during inference in the deep temporal correlation reinforcement network. For the graph isomorphic inference layer of the cross-period dependency extraction subnet, the graph node embedding vector sequence H_step_k is recorded for each inference step (a total of K steps), with each H_step_k having a shape of [N_nodes, D_emb]. For the frequency band self-attention layer of the signal component interaction subnet, the attention weight matrix sequence A_step_k is recorded for each inference step, with each A_step_k having a shape of [N_freq, N_freq]. The above records are summarized to obtain the internal inference state recording data State_rec, covering S different control scenarios.

[0167] Step S1220: Perform inference path visualization construction processing on the graph node embedding vector sequence, convert the graph node embedding vector in each inference step into node coordinates in two-dimensional space through dimensionality reduction mapping, and connect the node coordinates of adjacent inference steps through directed line segments to generate a graph isomorphic inference path trajectory graph.

[0168] In this embodiment, the graph node embedding vector sequences H_step_1, H_step_2, ..., H_step_K for each control scenario are visualized. The t-SNE dimensionality reduction algorithm is used to map the node embedding vectors in each H_step_k from D_emb dimension to a two-dimensional plane, obtaining the node coordinate set P_k. The coordinates of nodes with the same index between adjacent inference steps are connected by directed line segments (from P_k to P_{k+1}), resulting in the graph isomorphic inference path trajectory diagram Traj_gcn. This graph isomorphic inference path trajectory diagram illustrates the embedding evolution path of graph nodes during the inference process.

[0169] Step S1230: Perform attention focus migration analysis on the attention weight matrix sequence, extract the position of the maximum weight frequency point of the attention weight matrix in each inference step, and connect the positions of the maximum weight frequency points of adjacent inference steps through migration vectors to generate an attention focus migration trajectory map.

[0170] In this embodiment, the attention weight matrix sequence A_step_1, A_step_2, ..., A_step_K for each control scenario is analyzed. For each A_step_k, the row and column indices of the maximum weight are found, denoted as (r_k, c_k), representing the attention focus position in that step. The focus positions of adjacent steps are connected by a transfer vector (v_k=(r_{k+1}-r_k, c_{k+1}-c_k)) to generate an attention focus transfer trajectory map Traj_attn.

[0171] Step S1240: Input the graph isomorphic reasoning path trajectory map and the attention focus migration trajectory map into the trajectory pattern encoding layer of the reasoning logic extraction network, and jointly encode the path topology pattern in the graph isomorphic reasoning path trajectory map and the migration direction pattern in the attention focus migration trajectory map to generate a reasoning trajectory pattern vector.

[0172] In this embodiment, a reasoning logic extraction network is constructed, which includes a trajectory pattern encoding layer and a decision logic distillation layer. The trajectory pattern encoding layer employs a two-stream Transformer encoder. The first stream encodes the isomorphic reasoning path trajectory map Traj_gcn, extracting the path topology pattern feature F_gcn. The second stream encodes the attention focus transfer trajectory map Traj_attn, extracting the transfer direction pattern feature F_attn. F_gcn and F_attn are concatenated and mapped to the reasoning trajectory pattern vector V_mode through a fully connected layer.

[0173] Step S1250: Input the inference trajectory pattern vector into the decision logic distillation layer of the inference logic extraction network, and map the inference trajectory pattern vector into an interpretable set of decision logic rules. The set of decision logic rules consists of a rule chain composed of conditional predicates and conclusion predicates.

[0174] In this embodiment, the decision logic distillation layer employs a rule extraction-based algorithm. First, K-means clustering is used to divide V_mode into C clusters, each corresponding to a type of decision behavior. For each cluster, the inference steps corresponding to the vectors near the cluster center are extracted, and the correlation between its input features (such as current graph node embeddings and attention weights) and output actions (control parameter adjustments) is analyzed. A decision tree algorithm is then used to extract rule chains in the form IF (conditional predicate) THEN (conclusion predicate) from these data. The rules extracted from all clusters are merged to obtain an interpretable set of decision logic rules.

[0175] Step S1260: Perform rule coverage evaluation processing on each decision logic rule in the decision logic rule set. Calculate the frequency ratio of activation of each decision logic rule in different control scenarios of the internal inference state record data to generate a rule coverage score. Sort and filter all decision logic rules according to the rule coverage score, and retain decision logic rules with rule coverage scores exceeding a preset coverage threshold to form a core decision logic rule set.

[0176] In this embodiment, for each rule `Rule_i` in `Rules`, all inference steps of all control scenarios in `State_rec` are traversed, and the number of times the conditional predicate of the rule is satisfied (`N_act`) and the total number of inference steps (`N_total`) are counted. The rule coverage score `Cov_i` = `N_act` / `N_total`. A coverage threshold `Th_cov` is set. Rules with `Cov_i` > `Th_cov` are retained to form the core decision logic rule set `Rules_core`.

[0177] Step S1270: Compile the core decision logic rule set into a logical regularization constraint term that can be embedded in the deep temporal association reinforcement network. The logical regularization constraint term applies a distribution constraint that conforms to the core decision logic rule to the output of the joint inference layer when the deep temporal association reinforcement network performs joint inference operations.

[0178] In this embodiment, each rule in Rules_core is compiled into a logical regularization constraint term. For the rule IF(P)THEN(Q), a constraint loss function L_logic=max(0, f(P)-f(Q)) is constructed, where f(P) is the degree of satisfaction of the condition predicate, and f(Q) is the degree of satisfaction of the conclusion predicate. The constraint loss functions of all rules are weighted and summed to obtain the total logical regularization constraint term L_reg=∑w_i*L_logic_i.

[0179] Step S1280: In subsequent control cycles, when the joint inference layer of the deep temporal association reinforcement network generates the control parameter adjustment vector, the logical regularization constraint term is added to the loss function, so that the control parameter adjustment vector output by the joint inference layer simultaneously satisfies the data-driven inference result and the constraint conditions of the core decision logic rule set.

[0180] In this embodiment, during the training and inference process of the deep temporal association reinforcement network, the total loss function is set to L_total = L_data + λ * L_reg, where L_data is the original data-driven loss (such as mean squared error) and λ is the regularization coefficient. When the joint inference layer outputs the control parameter adjustment vector U, L_reg will penalize the output that violates the core decision logic rules during backpropagation, guiding the network to output an adjustment scheme that satisfies the logical constraints.

[0181] Step S1290: Generate a control parameter adjustment scheme by continuously constraining the output of the joint inference layer through the logical regularization constraint terms.

[0182] In this embodiment, during the inference process of each control cycle, the joint inference layer calculates L_reg and adds it to the total loss when generating the control parameter adjustment vector U. Gradient descent is used to continuously optimize the network weights, causing the network output to gradually converge to a solution space that simultaneously satisfies data fitting and logical rules. The final output U_causal, after multiple rounds of constraints, can make precise adjustments based on the measured signal data and conforms to the interpretable decision-making logic rules extracted from historical inference trajectories, generating a final, reliable, and interpretable control parameter adjustment scheme.

[0183] Based on the same inventive concept, please refer to Figure 3 This diagram illustrates a schematic block diagram of a deep learning-based intelligent instrument control system provided in an embodiment of this application. The system includes a central processing unit (CPU), a system memory comprising random access memory (RAM) and read-only memory (ROM), and a system bus connecting the system memory and the CPU. The deep learning-based intelligent instrument control system also includes a basic input / output system to facilitate information transfer between various devices within the computer, and a large-capacity storage device for storing the operating system, applications, and other program modules.

[0184] A basic input / output system includes a display for showing information and input devices such as a mouse and keyboard for user input. Both the display and the input devices are connected to the central processing unit via an input / output controller connected to the system bus. The basic input / output system may also include an input / output controller for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input / output controller also provides output to a display screen, printer, or other types of output devices.

[0185] A mass storage device is connected to the central processing unit via a mass storage controller connected to the system bus. The mass storage device and its associated computer-readable medium provide non-volatile storage for deep learning-based intelligent instrument control. Without loss of generality, computer-readable media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. According to various embodiments of this application, deep learning-based intelligent instrument control can also be connected to a remote computer on a network, such as the Internet. That is, deep learning-based intelligent instrument control can be connected to a network via a network interface unit connected to the system bus, or it can use a network interface unit to connect to other types of networks or remote computer systems.

[0186] In addition, in the specific embodiments of this application, data such as user information are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0187] The above are merely exemplary embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application shall be included within the protection scope of this application.

Claims

1. A deep learning-based intelligent instrument control method, characterized in that, The method includes: Acquire the raw measurement signal data collected by the smart instrument during the current measurement cycle and the historical measurement signal data corresponding to the previous control cycle stored in the smart instrument; The original measurement signal data is subjected to joint time-frequency domain decomposition processing to generate measurement signal decomposition features containing time-domain component sequences and frequency-domain component sequences. The historical measurement signal data is subjected to joint time-frequency domain decomposition processing to generate historical measurement signal decomposition features, which contain historical time-domain component sequences and historical frequency-domain component sequences. The measured signal decomposition features and the historical measured signal decomposition features are jointly input into a pre-constructed deep temporal correlation enhancement network. Joint inference operations are performed through the cross-period dependency extraction subnet and signal component interaction subnet within the deep temporal correlation enhancement network to generate a control parameter adjustment scheme. A set of instrument control commands is generated according to the control parameter adjustment scheme, and the set of instrument control commands is sent to the actuator and signal acquisition module through the control interface of the smart instrument. The set of instrument control commands includes actuator control voltage commands and signal acquisition module parameter commands.

2. The intelligent instrument control method based on deep learning according to claim 1, characterized in that, The process of performing time-frequency domain joint decomposition on the original measurement signal data to generate measurement signal decomposition features containing time-domain component sequences and frequency-domain component sequences, and performing time-frequency domain joint decomposition on the historical measurement signal data to generate historical measurement signal decomposition features, includes: The original measurement signal data is subjected to variable-scale time-domain window segmentation to obtain a sequence of signal segments with temporal adjacency. Local oscillation mode extraction is performed on each signal segment in the sequence to generate a set of narrowband oscillation components corresponding to each signal segment. The set of narrowband oscillation components consists of eigenmode components with a single dominant frequency oscillation attribute. For each signal segment, the intrinsic mode component with the largest change in zero crossover rate is selected from the narrowband oscillation component set as the dominant time-domain component of the signal segment, and the dominant time-domain components of all signal segments are spliced ​​together in chronological order to generate the time-domain component sequence. Perform Hilbert spectral mapping on each intrinsic mode component in the narrowband oscillation component set to obtain the instantaneous frequency change curve and instantaneous energy decay curve of each intrinsic mode component; Align the instantaneous frequency change curves of all intrinsic mode components within each signal segment along the time axis to generate the instantaneous frequency spectrum of that signal segment, and then stitch together the instantaneous frequency spectra of all signal segments in chronological order to generate the frequency domain component sequence. The time-domain component sequence is subjected to recursive graph generation processing. Based on the phase space reconstruction, the distance relationship between any two time points in the time-domain component sequence is calculated, and a time-domain recursive topology graph expressed in the form of a recursive matrix is ​​generated. The frequency domain component sequence is subjected to frequency band energy entropy extraction processing. The instantaneous energy decay curves of each intrinsic mode component in the frequency domain component sequence are accumulated by frequency point to generate a frequency point energy distribution curve. Based on the frequency point energy distribution curve, the proportion coefficient of each frequency point energy to the total energy is calculated, and a frequency band energy entropy vector is generated by frequency point sequence. The time-domain recursive topology diagram and the frequency band energy entropy vector are respectively used as the time-domain component and frequency-domain component of the measurement signal decomposition feature. The measurement signal decomposition feature contains independent time-domain components and frequency-domain components. The time-domain component of the measurement signal decomposition feature corresponds to the time-domain recursive topology diagram, and the frequency-domain component of the measurement signal decomposition feature corresponds to the frequency band energy entropy vector. Using the same processing flow, the historical measurement signal data is sequentially processed by variable-scale time-domain window segmentation, local oscillation mode extraction, time-domain dominant component selection, Hilbert spectrum mapping, instantaneous frequency spectrum generation, recursive graph generation, and frequency band energy entropy extraction to generate historical measurement signal decomposition features. The historical measurement signal decomposition features contain independent historical time-domain components and historical frequency-domain components, wherein the historical time-domain components correspond to the historical time-domain recursive topology graph, and the historical frequency-domain components correspond to the historical frequency band energy entropy vector.

3. The intelligent instrument control method based on deep learning according to claim 1, characterized in that, The process involves inputting the decomposition features of the measured signal and the decomposition features of the historical measured signal into a pre-constructed deep temporal correlation enhancement network. Joint inference operations are then performed via the cross-period dependency extraction subnet and the signal component interaction subnet within the deep temporal correlation enhancement network to generate a control parameter adjustment scheme, including: In the deep temporal correlation enhancement network, the temporal recursive topology graph contained in the measurement signal decomposition features is input to the graph isomorphic inference layer of the cross-period dependency extraction subnet. The neighborhood structure of each node in the temporal recursive topology graph is traversed through a message passing mechanism to generate a graph isomorphic representation vector. The historical temporal recursive topology graph contained in the historical measurement signal decomposition features is input to the graph isomorphic inference layer of the cross-period dependency extraction subnet. The neighborhood structure of each node in the historical temporal recursive topology graph is traversed through a message passing mechanism to generate a historical graph isomorphic representation vector. The graph isomorphic representation vector and the historical graph isomorphic representation vector are jointly input into the cross-period attention aggregation layer of the cross-period dependency extraction subnet. The periodic correlation matrix between the graph isomorphic representation vector and the historical graph isomorphic representation vector is calculated. The historical graph isomorphic representation vector is then weighted and aggregated using the periodic correlation matrix to generate a cross-period temporal dependency vector. The frequency band energy entropy vector contained in the decomposition features of the measured signal is input to the frequency band relationship coding layer of the signal component interaction subnet, and the relative position embedding coding of each frequency point position in the frequency band energy entropy vector is performed to generate the frequency band position coding vector. The frequency band energy entropy vector and the frequency band position encoding vector are input into the frequency band self-attention layer of the signal component interaction subnet, the frequency point association weights between each frequency point are calculated, and the frequency band energy entropy vector is weighted and summed according to the frequency point association weights to generate a frequency band context aggregation vector. The historical frequency band energy entropy vector and the frequency band position encoding vector contained in the historical measurement signal decomposition features are input into the frequency band self-attention layer of the signal component interaction subnet. The historical frequency point association weights between historical frequency points are calculated, and the historical frequency band energy entropy vector is weighted and summed according to the historical frequency point association weights to generate a historical frequency band context aggregation vector. The frequency band context aggregation vector and the historical frequency band context aggregation vector are jointly input into the cross-cycle frequency band fusion layer of the signal component interaction subnet. The interaction change between the frequency band context aggregation vector and the historical frequency band context aggregation vector at the same frequency point is calculated to generate a frequency band evolution difference vector. The cross-period temporal dependency vector and the frequency band evolution difference vector are input into the joint inference layer of the deep temporal correlation enhancement network. The cross-period temporal dependency vector and the frequency band evolution difference vector are fused into a joint inference input vector. The joint inference input vector is subjected to layer-by-layer dimensional transformation processing through the fully connected computing unit of the joint inference layer to generate a control parameter adjustment vector. Each element in the control parameter adjustment vector corresponds to the numerical adjustment amount of the control parameter. The control parameter adjustment vector is added element-wise to the current control parameters of the smart instrument in the current control cycle to generate an adjusted control parameter set, and the adjusted control parameter set is used as the control parameter adjustment scheme.

4. The intelligent instrument control method based on deep learning according to claim 3, characterized in that, After performing time-frequency domain joint decomposition processing on the original measurement signal data to generate measurement signal decomposition features containing time-domain component sequences and frequency-domain component sequences, and performing time-frequency domain joint decomposition processing on the historical measurement signal data to generate historical measurement signal decomposition features, the method further includes: According to the control parameter adjustment scheme, the actuator control voltage and signal acquisition module parameters of the intelligent instrument are updated, and the corrected measurement signal data of the next measurement cycle is obtained under the drive of the updated control parameters. The corrected measurement signal data is subjected to time-frequency domain joint decomposition processing to generate corrected measurement signal decomposition features. The corrected measurement signal decomposition features include a corrected time-domain recursive topology diagram and a corrected frequency band energy entropy vector. The frequency band energy entropy vector in the measured signal decomposition feature and the modified frequency band energy entropy vector in the modified measured signal decomposition feature are input to the frequency band response change evaluation module. The frequency band response change vector is generated by calculating the energy entropy change at the same frequency point. Correlation analysis is performed on the energy entropy change of each control parameter in the adjusted control parameter set included in the control parameter adjustment scheme and the corresponding frequency point in the frequency band response change vector to generate a parameter frequency band response correlation degree sequence. Based on the sorting result of each correlation degree value in the parameter frequency band response correlation degree sequence, a predetermined number of control parameters with the highest correlation degree are selected as the target sensitive parameter set. Linear mapping is performed on the control parameters in the target sensitive parameter set and the frequency band response change vector to generate a parameter response mapping relationship model. A parameter response correction branch is added to the joint inference layer of the deep temporal correlation enhancement network, and the target sensitive parameter set and the parameter response mapping relationship model are loaded into the parameter response correction branch. The parameter response correction branch performs secondary adjustment processing on the elements belonging to the target sensitive parameter set in the control parameter adjustment vector generated by the joint inference layer based on the mapping relationship, generating a corrected control parameter adjustment vector. The corrected control parameter adjustment vector replaces the original control parameter adjustment vector as the updated control parameter adjustment scheme, and the updated control parameter adjustment scheme drives the next control cycle of the smart instrument.

5. The intelligent instrument control method based on deep learning according to claim 1, characterized in that, The method further includes: The system acquires records of control parameter adjustment schemes generated by the smart instrument in multiple consecutive control cycles, forming a control parameter adjustment history sequence. The system performs parameter decoupling processing on the control parameter adjustment schemes corresponding to each control cycle in the control parameter adjustment history sequence, and extracts the actuator control voltage adjustment amount and signal acquisition module parameter adjustment amount corresponding to each control cycle. Arrange the actuator control voltage adjustment amount in the continuous control cycle in chronological order to generate the actuator adjustment time series, and arrange the signal acquisition module parameter adjustment amount in the continuous control cycle in chronological order to generate the acquisition module parameter adjustment time series; The actuator adjustment time series is subjected to trend decomposition processing, and the actuator adjustment time series is decomposed into control trend component series and control fluctuation component series through a sliding window. Differential autoregressive analysis is performed on the control trend component sequence to generate an actuator control trend prediction model. The actuator control voltage adjustment trend value for the next control cycle is predicted using the actuator control trend prediction model. Conditional probability distribution modeling is performed on the control fluctuation component sequence to generate a control fluctuation distribution model. The actuator control voltage adjustment fluctuation range for the next control cycle is predicted using the control fluctuation distribution model. The actuator control voltage adjustment trend value and the actuator control voltage adjustment fluctuation range are synthesized to generate an actuator control voltage adjustment prediction interval. When generating the control parameter adjustment scheme for the next control cycle, the actuator control voltage adjustment prediction interval is used as a constraint condition and input to the joint inference layer of the deep temporal correlation reinforcement network. The joint inference layer searches for a control parameter adjustment vector that meets the optimization index within the actuator control voltage adjustment prediction interval. By introducing the actuator control voltage adjustment prediction range as a constraint, the search space for control parameters of the deep temporal correlation reinforcement network in the next control cycle is reduced.

6. The intelligent instrument control method based on deep learning according to claim 3, characterized in that, The method further includes: Acquire environmental sensing data of the target site environment where the smart instrument is located, including temperature distribution data and vibration spectrum data of the target site; Spatial temperature field interpolation processing is performed on the temperature distribution data to generate a temperature field distribution map expressed in the form of a spatial grid, wherein each grid node of the temperature field distribution map corresponds to a temperature value. The vibration spectrum data is processed by extracting the dominant frequency components. A predetermined number of vibration dominant frequency components with the highest amplitude are extracted from the vibration spectrum data through amplitude spectrum analysis to generate a set of vibration dominant frequency components. The temperature field distribution map is processed by calculating the temperature difference between adjacent spatial grid nodes to generate a temperature gradient vector field. Each vector in the temperature gradient vector field represents the direction and rate of temperature change at the corresponding location. The temperature gradient vector field and the set of vibration dominant frequency components are input into the environmental state embedding coding network. The temperature gradient vector field and the set of vibration dominant frequency components are mapped to the environmental embedding vector space of the same dimension through a fully connected embedding layer to generate an environmental state embedding vector. The environmental state embedding vector is input into the environment adaptation branch of the deep temporal correlation enhancement network. The environmental state embedding vector is subjected to multi-layer nonlinear transformation processing to generate an environment adaptation modulation vector. In the joint inference layer of the deep temporal correlation enhancement network, the environment adaptation modulation vector, the cross-cycle temporal dependency vector, and the frequency band evolution difference vector are subjected to gated fusion processing. The corresponding elements of the cross-cycle temporal dependency vector and the frequency band evolution difference vector are weighted and modulated by each element of the environment adaptation modulation vector. The cross-cycle temporal dependency vector and the frequency band evolution difference vector modulated by the environment adaptation modulation vector are input to the fully connected computing unit of the joint inference layer to generate a control parameter adjustment vector that integrates environmental factors.

7. The intelligent instrument control method based on deep learning according to claim 1, characterized in that, The method further includes: The control effect feedback data generated by the smart instrument after executing the instrument control command set is obtained. The control effect feedback data includes the time series of the deviation between the actual measured value of the controlled variable and the target set value. The deviation time series is subjected to deviation pattern encoding processing, and the deviation time series is divided into positive deviation subsequences and negative deviation subsequences according to the deviation direction. The deviation amplitude distribution statistical processing is performed on the positive deviation subsequences and the negative deviation subsequences respectively to generate positive deviation distribution histograms and negative deviation distribution histograms. The distribution shape parameter extraction processing is performed on the positive deviation distribution histograms and the negative deviation distribution histograms respectively to generate a deviation distribution shape parameter set, which includes the skewness parameter and the kurtosis parameter of the deviation distribution. The deviation time series is input into the deviation feature extraction layer of the feedback strategy generation network, and temporal convolution processing is performed on the deviation time series to generate a deviation temporal feature vector. The deviation distribution morphological parameter set and the deviation temporal feature vector are input into the policy mapping layer of the feedback policy generation network. The deviation distribution morphological parameter set and the deviation temporal feature vector are concatenated and fused to generate a fused deviation feature vector. The policy classifier of the policy mapping layer performs policy category prediction processing on the fused deviation feature vector to generate a feedback policy category identifier. The feedback policy category identifier indicates the correction direction of the control parameter adjustment. The policy parameter generation layer of the feedback policy generation network generates a policy correction parameter vector by taking the fusion bias feature vector and the feedback policy category identifier as input. Each element in the policy correction parameter vector corresponds to the correction magnitude of the control parameter adjustment scheme. The strategy correction parameter vector and the control parameter adjustment scheme generated by the deep temporal correlation reinforcement network in the current control cycle are fused element by element to generate a control parameter adjustment scheme after feedback correction.

8. The intelligent instrument control method based on deep learning according to claim 1, characterized in that, The method further includes: The system acquires a set of multi-sample measurement signal data collected by the signal acquisition module of the smart instrument under different sampling configurations. It performs time-frequency domain joint decomposition processing on each sample in the set of multi-sample measurement signal data to generate a sample time-domain recursive topology diagram and a sample frequency band energy entropy vector corresponding to each sample. Each sample in the set of multi-sample measurement signal data corresponds to a sampling configuration. The time-domain recursive topology diagrams of samples corresponding to multiple consecutive control cycles under the same sampling configuration are input into the topology stability analysis layer of the sampling effect evaluation network. The graph structure similarity between the time-domain recursive topology diagrams of samples in adjacent cycles is calculated, and a topology stability evaluation value is generated. The frequency band energy entropy vectors of samples corresponding to multiple consecutive control cycles under the same sampling configuration are input into the frequency band coverage analysis layer of the sampling effect evaluation network. The energy distribution coverage ratio of the sample frequency band energy entropy vectors on the entire frequency band is calculated, and a frequency band coverage evaluation value is generated. The topology stability evaluation value and the frequency band coverage evaluation value are weighted and summed to generate a comprehensive sampling effect score. The comprehensive sampling effect scores corresponding to all sampling configurations are sorted and compared. The sampling configuration with the best comprehensive sampling effect score is selected from all sampling configurations as the optimal sampling configuration parameter. The optimal sampling configuration parameter is written into the configuration register of the signal acquisition module of the smart instrument, so that the signal acquisition module performs signal acquisition according to the optimal sampling configuration parameter in subsequent measurement cycles. After each generation of the control parameter adjustment scheme, the optimal sampling configuration parameter is compared with the signal acquisition module parameter instruction in the control parameter adjustment scheme. If the signal acquisition module parameter instruction in the control parameter adjustment scheme is inconsistent with the optimal sampling configuration parameter, the optimal sampling configuration parameter is used to replace the signal acquisition module parameter instruction.

9. The intelligent instrument control method based on deep learning according to claim 1, characterized in that, The method further includes: After the deep temporal correlation reinforcement network performs multiple rounds of control cycle inference, the control performance log data of the smart instrument is obtained. The control performance log data includes the deviation sequence between the actual value and the target value of the controlled variable in each control cycle and the adjustment step sequence of the control parameter adjustment scheme. The deviation sequence is subjected to performance degradation trend extraction processing. The slope of the residual change of the deviation sequence in continuous period is calculated by sliding linear fitting to generate a performance degradation slope sequence. The step size distribution statistical processing is performed on the adjustment step size sequence to calculate the distribution variance and distribution mean of the adjustment step size in continuous period to generate step size distribution statistical features. Joint performance diagnostic processing is performed on the performance degradation slope sequence and the step size distribution statistical features, and the performance degradation slope sequence and the step size distribution statistical features are mapped to performance degradation type identifiers through a preset performance diagnostic mapping function; Based on the performance degradation type identifier, a corresponding network structure parameter adjustment strategy is selected from the pre-built deep temporal correlation reinforcement network parameter library. The network structure parameter adjustment strategy includes an adjustment instruction for the number of graph isomorphic inference layer network layers of the cross-period dependency extraction subnet in the deep temporal correlation reinforcement network and an adjustment instruction for the number of attention heads of the frequency band self-attention layer of the signal component interaction subnet. The instructions for adjusting the number of layers in the graph isomorphic inference layer and the instructions for adjusting the number of attention heads in the frequency band self-attention layer are written into the network configuration file of the deep temporal correlation reinforcement network. This triggers the deep temporal correlation reinforcement network to perform a hot update operation on the structural parameters. After the hot update operation is completed, the original measurement signal data is re-processed by joint time-frequency domain decomposition to generate a hot-updated time-domain recursive topology graph and a hot-updated frequency band energy entropy vector. These are then input into the hot-updated deep temporal correlation reinforcement network to generate a hot-updated control parameter adjustment scheme.

10. A deep learning-based intelligent instrument control system, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the deep learning-based smart instrument control method according to any one of claims 1 to 9 by executing the machine-executable instructions.

Images