A script automatic generation method based on a pipelined neural network architecture

Abstract

The application relates to the field of artificial intelligence computing graph optimization, and discloses a script automatic generation method based on a pipelined neural network architecture, which comprises the following steps: converting a business logic directed topology flow data into a high-dimensional feature tensor with a logic dependence pointer; determining a time sequence correlation dispersion parameter according to statistical variance and a normalized difference ratio result; when the time sequence correlation dispersion parameter is greater than 0.85, freezing external weight input; a weight adjustment matrix covers a dynamic weight matrix in situ; a computing graph topology structure is calculated by decoupling a nonlinear Hadamard product feature; a data flow topology is rearranged by a script syntax control operator, and a script code is generated; a bidirectional feedback closed loop is constructed at a bottom operator level; an interlayer time delay of a layered reasoning layer is eliminated in situ; time sequence mismatch cascade amplification is blocked; logic deadlock and process suspension faults are avoided; and the real-time performance and data throughput of a computing pipeline are enhanced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence computational graph optimization technology, specifically to a method for automatic script generation based on a pipelined neural network architecture. Background Technology

[0002] Currently, in the deployment of large-scale dedicated computing models, the mainstream strategy adopts a cascaded neural network architecture to process temporal logic. It uses a hierarchical inference network to complete the feature topology transformation. The propagation alignment state of the feature tensor between different layers of the network directly determines the convergence stability of the computation graph. Under standard operating conditions, the operator unit follows the topological order to perform hierarchical feedforward computation to produce the control feature state flow. In order to ensure the continuity of the feedforward feature flow, the cascaded architecture is usually configured with a cache queue to balance the throughput rate of each level of operator.

[0003] However, such architectures are prone to asynchronous disturbances under non-ideal operating conditions during forward flow. Due to the lack of parameter adaptive constraint paths, when timing mismatch occurs between the front-end network and the intermediate pipeline, multi-source inputs cause high-frequency phase slippage in action nodes. This leads to the architecture introducing static blocking waiting time to maintain topological order. This inertia causes long-term implicit phase delays in feature flow and disrupts steady-state relationships, resulting in degraded resource utilization efficiency in the sequence generation stage. The aforementioned limitations of hardware and cascaded architecture deployed for specific computing models cannot fundamentally eliminate timing mismatches. The automated script generation control methods at the software layer also have shortcomings. For example, Chinese invention patent application CN119127721A discloses an AI-based... The Selenium automated script generation method and system for the algorithm establishes a dataset by cleaning and labeling automated scripts, and outputs automated scripts using the Transformer model combined with the positive and negative feedback mechanism of reinforcement learning. The underlying attribute of this technical solution is that it relies on static preset premises in the offline state. The model structure parameters are only set and adjusted according to the static indicators of the initial test task. It lacks online perception and adaptive path to cope with dynamic asynchronous disturbances during runtime. When facing non-ideal working conditions such as high-frequency interaction of multi-source heterogeneous interfaces or deep logic nesting, the static software control mechanism cannot offset the implicit phase delay transmitted between layers in real time. It is easy to trigger the manifold divergence of the computation graph due to the cascading amplification of timing mismatch, which can cause logical deadlock or process suspension failure in the system.

[0004] Therefore, the technical problem to be solved by this invention is how to perceive the temporal dependency dispersion in hierarchical inference online and construct a dynamic causal closed loop that aligns parameter allocation with temporal sequence, so as to offset the unidirectional transmission resistance of the cascaded architecture while constraining the stable convergence of the computation graph manifold. Summary of the Invention

[0005] To address the problems in the background art, this invention proposes an automatic script generation method based on a pipelined neural network architecture, comprising the following steps: Step S1: Obtain the business logic directed topology flow data and the branch prediction probability of multi-level action nodes. Use the front-end neural network to map the business logic directed topology flow data into a high-dimensional feature tensor containing a logical dependency pointer. Extract the variance features of the high-dimensional feature tensor within the sliding time window. Multiply the branch prediction probability of the multi-level action nodes with the logical dependency pointer and calculate the normalized difference comparison value. Output the dimensionless temporal correlation dispersion parameter. Step S2: In the hierarchical feedforward calculation of the multi-level neural network, the data transmission delay status between layers is determined. When the dimensionless temporal correlation discreteness parameter is greater than the safety critical threshold of 0.85, the data feedforward decoupling and reconstruction process is triggered. The external weight input channel is disconnected, and the dynamic weight allocation matrix in the previous neural network is covered and modified in situ using the weight adjustment matrix. The in situ adjustment increment matrix in the weight adjustment matrix is ​​calculated and nonlinear Hadamard product is obtained with the high-dimensional feature tensor. The topological structure of the computation graph is constrained to the topological space. Step S3: Receive the reconstructed high-dimensional feature tensor, configure the script syntax control operator to complete the data flow topology rearrangement, determine the convergence point weight adjustment amount using matrix linear mapping rules when determining the priority of control instruction generation, and finally generate test script code.

[0006] Preferably, the determination of the inter-layer data transmission delay status in the hierarchical feedforward calculation of the multi-level neural network in step S2 includes the following steps: Step S21, when the hierarchical inference delay is greater than the critical delay threshold of 5ms, the dimensionality reduction hedging process is started; Step S22, the temporal phase difference is quantized into a delay compensation weight increment matrix, the constant coefficient steady-state reference tensor saved in the historical data section is called to dynamically allocate the weight matrix, the static topological constraint matrix is ​​used to complete the feature slicing of the high-dimensional feature tensor, and the discretized gradient is used to eliminate the inter-layer cumulative delay.

[0007] Preferably, step S1 includes the following sub-steps: step S11, obtaining the branch prediction probability of the multi-level action node; step S12, performing normalized difference comparison on the product of the branch prediction probability of the multi-level action node and the logical dependency index, and outputting the dimensionless temporal correlation dispersion parameter.

[0008] Preferably, step S2, which puts the external weight input into a frozen state, includes the following sub-steps: step S23, after receiving the in-situ adjustment increment matrix, disconnecting the channel of the external weight input; step S24, locking the update step size of the external weight input to 0, and controlling the weight update of the preceding neural network to respond independently to the in-situ adjustment increment matrix.

[0009] Preferably, step S2, which involves calculating the nonlinear Hadamard product based on the in-situ adjustment increment matrix and the high-dimensional feature tensor and constraining the computation graph topology to the topological space, includes the following sub-steps: Step S25, calculating the element-wise product of the in-situ adjustment increment matrix and the high-dimensional feature tensor, and outputting the decoupled core feature matrix; Step S26, performing orthogonal projection on the eigenvalues ​​of the core feature matrix using boundary constraints, and constraining the computation graph topology to the topological space.

[0010] Preferably, step S22, which involves slicing the high-dimensional feature tensor using a static topological constraint matrix and eliminating inter-layer cumulative delay using discretized gradients, includes the following sub-steps: Step S221, using a static topological constraint matrix to intercept the topological unfolding of the high-dimensional feature tensor during forward propagation; Step S222, slicing the high-dimensional feature tensor along the time axis, calculating the differential gradient between adjacent slices, and eliminating inter-layer data transmission delay using the differential gradient.

[0011] Preferably, the configuration of the script syntax control operator in step S3 to complete the data flow topology reordering includes the following sub-steps: Step S31, constructing a temporal dependency graph of the script syntax based on the reconstructed high-dimensional feature tensor; Step S32, using the script syntax control operator to complete the spatial arrangement adjustment and time axis alignment of the control flow nodes in the temporal dependency graph.

[0012] Preferably, step S3, in determining the convergence point weight adjustment amount using the matrix linear mapping rule when determining the control instruction generation priority, includes the following sub-steps: Step S33, extract the instruction type and control depth of the script syntax control operator, and construct a linear evaluation index for priority; Step S34, use the linear evaluation index to sort the priority of each control instruction, and calculate the convergence point weight adjustment amount in the time-series dependency graph.

[0013] Preferably, generating test script code includes the following sub-steps: Step S35, applying the convergence point weight adjustment to each control flow node in the time-series dependency graph; Step S36, outputting the test script control instruction flow according to each control flow node after spatial arrangement adjustment and time axis alignment.

[0014] Preferably, the constant-coefficient steady-state reference tensor stored in the historical data section is invoked. Its initialization deployment and solidification storage steps include: under the initialization deployment condition where no directed topology data has been entered, the test operator cluster is adjusted to an idle state, the system latency of the test host is obtained, and the dynamic weight allocation matrix of the initial state is read; by statistically analyzing the cumulative deviation of the feedforward calculation of each action layer, the row vectors of the mapping weights of neurons at each level are normalized and aligned, and the background noise features are ablated in situ; after the idle baseline alignment, the mean of the sampled steady-state phase space neuron state distribution tensor is calculated, and the calculated state probability steady-state matrix is ​​used as the constant-coefficient steady-state reference tensor and solidified and stored in the physical reference address of the non-volatile memory.

[0015] The beneficial effects of this invention are: 1. In the automatic script generation of the pipelined neural network architecture, the unit collects the directed topology flow data of the business logic and the physical latency jitter. The front-end network deconstructs the directed topology flow data into discrete logic node feature vectors and configures the dynamically allocated weight matrix. The intermediate pipeline state machine performs hierarchical feedforward calculations according to the period order of the feature vectors and produces operator operation feedback in real time. This feedback corrects the dynamically allocated weight matrix of the front-end network in situ, thereby constructing a bidirectional causal closed loop of data interweaving between parameter adjustment and state machine deduction. This closed loop mechanism offsets the inherent unidirectional transmission resistance of the cascaded architecture at the source of model feature flow, eliminates computational bubbles under long-distance temporal nesting constraints, and ensures stable convergence of the computation graph inference path manifold.

[0016] 2. The computation node monitors the intermediate feature tensor inside the preceding neural network online and extracts the statistical variance within a preset sliding time window. The branch prediction probability of the heterogeneous action node is multiplied by the logical dependency index and normalized difference is compared to determine the dimensionless temporal correlation dispersion parameter. When this parameter is greater than the critical threshold of 0.85, the data feedforward decoupling process is triggered, so that the external weight input enters a completely quiescent state. Based on the in-situ adjustment increment of the dynamically allocated weight matrix, nonlinear Hadamard product feature decoupling is performed with the high-dimensional feature tensor to constrain the computation graph manifold to a preset topological space, thus mitigating the risk of manifold divergence induced by high-frequency asynchronous disturbances and multi-level control flow nesting.

[0017] 3. When the intermediate pipeline state machine determines the hardware propagation delay state during hierarchical feedforward computation, if the hierarchical inference delay exceeds the set critical delay threshold of 5ms, the system immediately disconnects the current adaptive path and switches to the dimension reduction hedging mechanism. The timing phase difference is quantized into a logic rack bounce compensation parameter. The constant coefficient steady-state reference tensor stored in the historical data section is called to replace the current dynamically allocated weight matrix in situ. The high-dimensional feature tensor is statically hard sliced ​​through the static topological constraint matrix. The discretized gradient hedging method is used to smooth out the implicit phase delay caused by the accumulation of multi-level network feedback, and eliminate the specific computing system deadlock and process suspension failure caused by the cascade amplification of timing mismatch. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the control logic and topology reconstruction of an automatic script generation method according to the present invention. Figure 2 This is a diagram showing the overall architecture and module interaction of an automatic script generation system according to the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] A method for automatically generating scripts based on a pipelined neural network architecture includes the following steps: Step S1: Obtain the business logic directed topology flow data and the branch prediction probability of multi-level action nodes. Use the front-end neural network to map the business logic directed topology flow data into a high-dimensional feature tensor containing a logical dependency pointer. Extract the variance features of the high-dimensional feature tensor within the sliding time window. Multiply the branch prediction probability of the multi-level action nodes with the logical dependency pointer and calculate the normalized difference comparison value. Output the dimensionless temporal correlation dispersion parameter. Step S2: In the hierarchical feedforward calculation of the multi-level neural network, the data transmission delay status between layers is determined. When the dimensionless temporal correlation discreteness parameter is greater than the safety critical threshold of 0.85, the data feedforward decoupling and reconstruction process is triggered. The external weight input channel is disconnected, and the dynamic weight allocation matrix in the previous neural network is covered and modified in situ using the weight adjustment matrix. The in situ adjustment increment matrix in the weight adjustment matrix is ​​calculated and nonlinear Hadamard product is obtained with the high-dimensional feature tensor. The topological structure of the computation graph is constrained to the topological space. Step S3: Receive the reconstructed high-dimensional feature tensor, configure the script syntax control operator to complete the data flow topology rearrangement, determine the convergence point weight adjustment amount using matrix linear mapping rules when determining the priority of control instruction generation, and finally generate test script code.

[0022] Preferably, the determination of the inter-layer data transmission delay status in the hierarchical feedforward calculation of the multi-level neural network in step S2 includes the following steps: Step S21, when the hierarchical inference delay is greater than the critical delay threshold of 5ms, the dimensionality reduction hedging process is started; Step S22, the temporal phase difference is quantized into a delay compensation weight increment matrix, the constant coefficient steady-state reference tensor saved in the historical data section is called to dynamically allocate the weight matrix, the static topological constraint matrix is ​​used to complete the feature slicing of the high-dimensional feature tensor, and the discretized gradient is used to eliminate the inter-layer cumulative delay.

[0023] Preferably, step S1 includes the following sub-steps: step S11, obtaining the branch prediction probability of the multi-level action node; step S12, performing normalized difference comparison on the product of the branch prediction probability of the multi-level action node and the logical dependency index, and outputting the dimensionless temporal correlation dispersion parameter.

[0024] Preferably, step S2, which puts the external weight input into a frozen state, includes the following sub-steps: step S23, after receiving the in-situ adjustment increment matrix, disconnecting the channel of the external weight input; step S24, locking the update step size of the external weight input to 0, and controlling the weight update of the preceding neural network to respond independently to the in-situ adjustment increment matrix.

[0025] Preferably, step S2, which involves calculating the nonlinear Hadamard product based on the in-situ adjustment increment matrix and the high-dimensional feature tensor and constraining the computation graph topology to the topological space, includes the following sub-steps: Step S25, calculating the element-wise product of the in-situ adjustment increment matrix and the high-dimensional feature tensor, and outputting the decoupled core feature matrix; Step S26, performing orthogonal projection on the eigenvalues ​​of the core feature matrix using boundary constraints, and constraining the computation graph topology to the topological space.

[0026] Preferably, step S22, which involves slicing the high-dimensional feature tensor using a static topological constraint matrix and eliminating inter-layer cumulative delay using discretized gradients, includes the following sub-steps: Step S221, using a static topological constraint matrix to intercept the topological unfolding of the high-dimensional feature tensor during forward propagation; Step S222, slicing the high-dimensional feature tensor along the time axis, calculating the differential gradient between adjacent slices, and eliminating inter-layer data transmission delay using the differential gradient.

[0027] Preferably, the configuration of the script syntax control operator in step S3 to complete the data flow topology reordering includes the following sub-steps: Step S31, constructing a temporal dependency graph of the script syntax based on the reconstructed high-dimensional feature tensor; Step S32, using the script syntax control operator to complete the spatial arrangement adjustment and time axis alignment of the control flow nodes in the temporal dependency graph.

[0028] Preferably, step S3, in determining the convergence point weight adjustment amount using the matrix linear mapping rule when determining the control instruction generation priority, includes the following sub-steps: Step S33, extract the instruction type and control depth of the script syntax control operator, and construct a linear evaluation index for priority; Step S34, use the linear evaluation index to sort the priority of each control instruction, and calculate the convergence point weight adjustment amount in the time-series dependency graph.

[0029] Preferably, generating test script code includes the following sub-steps: Step S35, applying the convergence point weight adjustment to each control flow node in the time-series dependency graph; Step S36, outputting the test script control instruction flow according to each control flow node after spatial arrangement adjustment and time axis alignment.

[0030] Preferably, the constant-coefficient steady-state reference tensor stored in the historical data section is invoked. Its initialization deployment and solidification storage steps include: under the initialization deployment condition where no directed topology data has been entered, the test operator cluster is adjusted to an idle state, the system latency of the test host is obtained, and the dynamic weight allocation matrix of the initial state is read; by statistically analyzing the cumulative deviation of the feedforward calculation of each action layer, the row vectors of the mapping weights of neurons at each level are normalized and aligned, and the background noise features are ablated in situ; after the idle baseline alignment, the mean of the sampled steady-state phase space neuron state distribution tensor is calculated, and the calculated state probability steady-state matrix is ​​used as the constant-coefficient steady-state reference tensor and solidified and stored in the physical reference address of the non-volatile memory.

[0031] Example 1: When a current script automatic generation method based on a pipelined neural network architecture is deployed in the field of test script automatic generation for cross-platform automated testing of heterogeneous software systems, the general computing system needs to cope with the high-frequency, variable, and nested dependency test action timing conditions between cross-platform multi-source heterogeneous interfaces. When the nesting depth of the control flow logic of the software test script to be generated exceeds the critical threshold and the nesting layer of control loops and conditional branches exceeds 3 layers, the pipelined operators of the traditional cascaded neural network lack the ability to dynamically reverse the high-dimensional architecture of the global context of the preceding stage, resulting in propagation resistance within the computation graph. Feature diffusion generates computational bubbles in time sequencing, leading to continuous logical retreats. This, amplified by cascading delays from external asynchronous verification feedback, causes a temporal deadlock in the operator flow network within a specific computational model, resulting in frequent computational process suspensions. The computational system employing the method claimed in this invention, under these conditions, acquires initial input data from heterogeneous software systems automatically generated within the test script's domain through an acquisition unit. This initial input data includes high-dimensional business logic directed topology flow data as a source of functional constraints, and physical latency jitter measured by the system timer during the execution of test tasks by the underlying computational architecture. The neural network processing module receives directed topology flow data of business logic, maps it into a high-dimensional feature tensor containing logical dependency pointers, assigns a corresponding dynamically allocated weight matrix to the high-dimensional feature tensor, and in this state, compute nodes measure intermediate process indicators in parallel. It extracts the statistical variance of the high-dimensional feature tensor input to the current network operator layer within a preset sliding time window, calculates the product of the branch prediction probability and the logical dependency pointer for each action node in the high-dimensional feature tensor, performs normalized difference comparison on the product of the branch prediction probability and the logical dependency pointer to calculate the residual ratio of its deviation from the data feature topology distribution, and outputs a dimensionless value. The temporal correlation dispersion parameter is used to characterize the convergence state of the computation graph topology in real time. The central processing unit reads 128 topological sampling points within the sliding time window and calculates the statistical variance of the values ​​of each feature channel in the high-dimensional feature tensor. The signal processing unit retrieves the high-dimensional feature tensor of the current time step and calculates the multidimensional mean vector. It retrieves the pre-stored mean vector of the stationary phase space neuron state distribution tensor from the non-volatile memory as the reference mean vector. It calculates the absolute difference between the multidimensional mean vector and the reference mean vector on each feature channel axis. After taking the square root of the sum of squares of all absolute differences, it generates the feature space distribution residual.

[0032] The computational unit uses the residual of the feature space distribution as the numerator and divides it by the statistical variance as the denominator to obtain the temporal correlation dispersion parameter through division, which serves as an indicator for judging the divergence of the inter-layer topological manifold. Specifically, the directed topological flow data is transformed into a dense matrix according to the node sequence and edge relationship and input into the front-end neural network. The multi-layer fully connected layers and recurrent network units of the front-end neural network perform layer-by-layer aggregation and latent space mapping of node features, and output a feature tensor with a dimension of 128. In this mapping process, the front-end neural network has set up an independent position encoding attention mechanism channel in its last layer. It uses the topological distance between the source node and the destination node in the graph structure and the logical nesting depth to calculate discrete index weights. These index weights are concatenated at the end of the feature tensor in the form of multi-dimensional floating-point dense vectors, thereby implicitly embedding numerical logical dependency pointers that can indicate the depth of upstream and downstream dependencies of control nodes in a continuous high-dimensional numerical matrix, and finally forming a composite high-dimensional feature tensor containing information on the upstream and downstream dependencies of business logic.

[0033] The intermediate pipelined state machine receives high-dimensional feature tensors and determines the inter-layer data transmission delay status during hierarchical feedforward computation. When the dimensionless temporal correlation discreteness parameter is not greater than the safety threshold of 0.85, the intermediate pipelined state machine processes the hierarchical feedforward computation sequentially under standard operating conditions according to the order of feature vectors of each discrete logic node. However, when the dimensionless temporal correlation discreteness parameter is greater than the safety threshold of 0.85, the system forcibly triggers the data feedforward decoupling and reconstruction process. This is achieved by disconnecting the external weight input channel and locking the update step size of the external weight input to 0 to ensure that the external weight input... Entering a resting state, the dynamic weight allocation matrix in the preceding neural network is modified in situ using the weight adjustment matrix. When calculating the in-situ adjustment increment in the weight adjustment matrix, the prediction probability tensor of the current step test action is subtracted from the steady-state reference tensor saved by the historical cross-section power outage to obtain the prediction bias matrix. Then, the phase delay difference generated by the physical delay jitter in the forward computation graph is summed with a constant 1 to serve as the step denominator. The in-situ adjustment increment of the dynamic weight allocation matrix is ​​determined by dividing the prediction bias matrix by the step denominator, driving the preceding neural network processing module to adjust in situ. The incremental and high-dimensional feature tensors are subjected to element-wise Hadamard product to achieve feature decoupling, outputting a core feature matrix. Boundary constraints are used to orthogonally project the eigenvalues ​​of the core feature matrix, constraining the computation graph topology to a preset topological space. This creates a bidirectional dependency loop where the weight updates of the front-end neural network and the hierarchical feedforward computation of the intermediate pipeline state machine interfere with each other's underlying parameters. When the hierarchical inference delay of the intermediate pipeline state machine due to hardware transmission latency exceeds the set critical delay threshold of 5ms, the system immediately disconnects the current adaptive path and forcibly switches to a dimensionality reduction hedging process. This system utilizes… The timer collects the timing phase difference and quantizes it into a delay compensation weight increment matrix. It calls the constant coefficient steady-state reference tensor stored in the historical data section to constrain the dynamically allocated weight matrix and replace it in situ. It intercepts the topological expansion of the high-dimensional feature tensor in forward propagation through the static topological constraint matrix. It slices the high-dimensional feature tensor along the temporal direction and calculates the differential gradient between adjacent slices. It uses the discretized gradient hedging method composed of differential gradients to smooth the implicit phase delay caused by the accumulation of multi-level network feedback. It blocks the path of temporal mismatch cascade amplification at the internal operator level of the deep learning model.

[0034] The output processing module receives the high-dimensional feature tensor after the aforementioned causal collaborative update and decoupling reconstruction. It then invokes script syntax control operators to construct a script syntax temporal dependency graph. These operators are used to adjust the spatial arrangement and temporal sequence alignment of control flow nodes in the temporal dependency graph, achieving data flow topology rearrangement. When determining the priority of control command generation, the module extracts the command type and control depth of the script syntax control operators to construct a linear evaluation index for priority. This linear evaluation index is used to prioritize each control command, and the convergence point weight adjustment is determined based on matrix linear mapping rules. The convergence point weight adjustment tensor is applied to each control flow node in the temporal dependency graph. Based on the spatial arrangement adjustment and temporal sequence alignment of each control flow node, the test script control instruction flow is output to drive the downstream heterogeneous software system to respond to automated testing. This enables the neural network computation graph, which is prone to computational bubbles and process suspension due to long-distance temporal nested dependencies, to achieve stable convergence in linear time. It eliminates the temporal mismatch of specific computing systems under extreme high-frequency disturbance conditions and ensures the real-time state convergence and high-dimensional feature matrix data throughput of the computing pipeline when processing deeply nested logic.

[0035] Example 2: When verifying the stability and adaptability of the automatic test script generation method in cross-platform automated testing of heterogeneous software systems, the central processing unit runs in a computing cluster with a floating-point computing power of greater than or equal to 100 trillion floating-point operations per second. It receives topology instruction stream data generated by asynchronous interface calls and containing timing phase deviations. When the data acquisition unit acquires the directed topology stream data of high-dimensional business logic, it superimposes Gaussian white noise with a signal-to-noise ratio of 20dB into the branch prediction probability signal of multi-level action nodes. At the same time, it introduces random time delay jitter discretely distributed between 0.5ms and 4.5ms using a timer to construct a non-ideal working condition test scenario. When determining the number of sampling points for the preset sliding time window, a window that is too long causes smoothing dilution of high-frequency timing-dependent features, while a window that is too short causes divergence in the calculation of higher-order statistical variance. To balance the dilution of high-frequency timing-dependent features with the convergence of statistical variance calculation, the signal processing unit measures the bus characteristic bandwidth of the input directed topology instruction stream. When the bus characteristic bandwidth is in the range of 500 to 2000 calls per second, to meet the requirements of low-order cutoff frequency filtering and eliminate high-frequency feature aliasing, the central processing unit controls the length of the preset sliding time window to be between 3 and 5 times the current processing cycle time span, specifically for deeply nested discrete logic nodes. In a deep control loop with a depth of 4, the number of sampling points within the preset sliding time window is determined to be 128. In the test group using the method of this invention, the computing system receives high-dimensional business logic directed topology flow data. The front-end neural network processing module converts the high-dimensional business logic directed topology flow data into a high-dimensional feature tensor containing a logical dependency pointer. The computing unit extracts the statistical variance of the high-dimensional feature tensor within the preset sliding time window of 128 sampling points, and calculates the product of the branch prediction probability of each action node in the high-dimensional feature tensor and the logical dependency pointer. Then, the product of the branch prediction probability and the logical dependency pointer is normalized and differentially compared to calculate the deviation data. The residual ratio of the characteristic topology distribution is used to output the dimensionless temporal correlation discreteness parameter. When the input random delay jitter is under low-intensity disturbance condition with an average value of 1.23ms, the average measured value of the dimensionless temporal correlation discreteness parameter output by the calculation system is 0.412. The intermediate pipeline state machine completes hierarchical feedforward inference according to the order of the feature vectors of each discrete logic node. When performing multiplication, the system converts the logical dependency pointer, which is a discrete address feature, into its corresponding multidimensional numerical sparse vector through one-hot encoding, and then into a dependency strength evaluation scalar with continuous numerical attributes through the final mapping layer of the previous neural network.

[0036] By performing a one-to-one scalar multiplication operation between the continuous branch prediction probability floating-point number of each action node and the dependency strength evaluation scalar, the logical mismatch problem of discrete pointers being unable to directly perform mathematical multiplication with continuous probability scalars is eliminated. This allows the final product to accurately and quantitatively reflect the probabilistic temporal constraint strength of the action node. When the system is under high-intensity disturbance conditions with a temporal dependency control flow node nesting depth of 5 levels and random delay jitter averaging 3.87ms, the real-time measurement value of the dimensionless temporal correlation dispersion parameter reaches 0.8. 94. When the measured value exceeds the safety threshold of 0.85, the computing system closes the external weight input channel and locks the update step size of the external weight input to 0 to maintain the quiescent state of the external weight input. Simultaneously, the computing unit subtracts the constant-coefficient steady-state reference tensor stored in the historical data section from the current step's state prediction probability tensor to calculate the prediction deviation matrix. The signal processing unit calls a two-stage Butterworth low-pass filter to process the physical delay jitter data stream collected by the system timer, with a cutoff frequency set to 500Hz to filter out random high-frequency non-communication noise introduced by external operating system task scheduling and hardware interrupts. The central processing unit retrieves the floating-point value of the steady-state communication delay jitter retained after filtering, divides it by the processor core's standard clock cycle to calculate a dimensionless scalar, and uses the resulting dimensionless value as the phase delay difference. The random delay jitter is then summed with the phase delay difference generated in the forward computation graph and a constant 1 to establish the step denominator. The in-situ adjustment increment of the dynamically allocated weight matrix is ​​calculated by dividing the prediction deviation matrix by the step denominator, and the in-situ adjustment increment is calculated by the front-end neural network processing module. The nonlinear Hadamard product between the high-dimensional feature tensor and the core feature matrix is ​​generated. After orthogonally projecting the eigenvalues ​​of the core feature matrix through boundary constraints, the core feature matrix is ​​transmitted to the intermediate pipeline state machine. This makes the weight update of the front-end neural network and the hierarchical feedforward calculation of the intermediate pipeline state machine form a two-way dependent closed loop in which the underlying parameters interfere with each other. The single inference delay measurement value of the computation graph in 1000 consecutive high-order nested control flow script instruction generation loops is maintained at 4.12ms, and no timing deadlock faults or computation process suspension states occur.

[0037] When the intermediate pipeline state machine experiences hierarchical inference latency exceeding the critical latency threshold of 5ms due to hardware transmission delay, the computing system disconnects the adaptive path and switches to a dimensionality reduction hedging mode. The system timer collects the timing phase difference, and the central processing unit converts the timing phase difference into a delay compensation weight increment matrix. It then calls the constant coefficient steady-state reference tensor stored in the historical data section to replace the current dynamically allocated weight matrix in situ. The static topology constraint matrix intercepts the topological expansion of the high-dimensional feature tensor during forward feedforward graph propagation. The central processing unit slices the high-dimensional feature tensor along the timing direction and calculates the differential gradient between adjacent feature slices. The discretized gradient hedging structure formed by the differential gradient is used to smooth out the phase delay caused by the accumulation of multi-level network feedback. The graph operator architecture of the neural network model blocks the path of cascaded amplification of timing mismatch. In the actual conversion process, the quantization rule for converting the timing phase difference into a delay compensation weight increment matrix is ​​as follows: the central processing unit extracts the floating-point value of the timing phase difference and divides it by the standard clock cycle to complete the dimensionless scalarization.

[0038] Then, a base matrix of all ones with the same number of rows and columns as the dynamic weight allocation matrix of the previous neural network is constructed. The scalarized temporal phase difference value is used as the global step size multiplication coefficient and multiplied by this base matrix of all ones to distribute the value evenly across the parameter nodes of the matrix. Finally, combining the weight distribution characteristics of the constant coefficient steady-state reference tensor read from non-volatile memory, positional correlation attenuation adjustment is performed on each parameter node to generate a high-dimensional delay compensation weight increment matrix that can directly combat temporal deviations. The output processing module receives the high-dimensional feature tensor after causal collaborative update and decoupling reconstruction, calls the script syntax control operator to construct the script syntax temporal dependency graph, and uses the script syntax control operator to adjust the spatial arrangement of control flow nodes in the temporal dependency graph and align the temporal sequences to complete the data flow topology rearrangement. When determining the priority of control command generation, the output processing module extracts the command type and control depth of the script syntax control operator to construct a linear evaluation index of priority, and uses the linear evaluation index to rank each control command. Priority is determined by the CPU, which prioritizes each control instruction and constructs a diagonal matrix of instruction priorities with the same dimension as the number of control flow nodes. The diagonal values ​​are set as the reciprocals of the priority numbers of each control instruction, from highest to lowest priority. The CPU then calls a preset transformation matrix to multiply the instruction priority diagonal matrix with the 128-dimensional feature tensor output by the preceding neural network. The resulting product matrix serves as the convergence point weight adjustment tensor. Based on the matrix linear mapping rule, the convergence point weight adjustment tensor is determined and applied to each control flow node in the temporal dependency graph. After spatial arrangement adjustment and temporal sequence alignment, each control flow node outputs test script control instructions to drive the downstream heterogeneous software system to respond to automated testing. This ensures that the neural network computation graph, which suffers from computational bubbles and process suspensions due to long-distance temporal nested dependencies, converges stably in linear time. This eliminates temporal mismatches in the computing system under high-frequency disturbance conditions and maintains the real-time convergence of the computing pipeline and the data throughput of the high-dimensional feature matrix when processing deeply nested logic.

[0039] The test setup included a first control group, a second control group, and a third control group to establish a performance benchmark. The first control group disabled the in-situ overlay update mechanism of the weight adjustment matrix while retaining dynamic allocation from the preceding network. The second control group disabled feature decoupling processing based on the high-dimensional feature tensor Hadamard product to break the parameter coordination loop between the preceding neural network and the intermediate pipeline state machine. The third control group increased the activation reconstruction threshold from the safety threshold of 0.85 to 0.95. Under nested control conditions including Gaussian white noise and a random time delay jitter with an average value of 3.87ms, the first control group, lacking an in-situ error hedging operator, ... The cumulative phase delay caused by multi-level network feedback cannot be mitigated, and its forward inference delay measurement increases exponentially with the number of iterations, reaching 23.41ms in the 42nd iteration, resulting in computational deadlock due to amplified timing mismatch. The second control group, lacking feature decoupling orthogonal projection, experiences high-dimensional divergence in the forward-feedback computation graph manifold, distorting the data feature topology. Its script syntax timing dependency graph generation failure rate is 41.6%, creating a bottleneck in cascaded network propagation resistance. In the superboundary parameter test of the third control group, when the judgment threshold is set to 0.95, the system... The dimensionless temporal correlation discreteness parameter, within the nonlinear inflection point saturation region of 0.85 to 0.95, did not trigger reconstruction. Its internal inference delay growth rate exhibited a step-like distortion, and the phase space manifold curvature within the computation graph diverged. Reconstruction was only triggered after the temporal correlation discreteness parameter reached 0.954, at which point the inference delay parameter adjusted towards a stable state. The aforementioned nonlinear inflection point data confirms that the safety critical threshold of 0.85 serves as the convergence boundary control window for the computational model. Furthermore, regarding the selection of the judgment threshold, if the safety critical threshold is below 0.85, the system will frequently trigger feedforward solutions under low-intensity disturbances. Coupling the process and locking the external weight channel causes the network to lose its normal adaptive online learning ability, and the inference latency will increase due to the increased reconstruction overhead. If it is higher than 0.85, such as 0.95 in the test, the network has entered the nonlinear divergence saturation region, and the reconstruction mechanism cannot play a role before the timing mismatch is amplified, and the system will directly fall into deadlock. Similarly, if the critical latency threshold is lower than 5ms, it is easy to accidentally trigger the dimensionality reduction hedging mode, which will destroy the coherence of the high-dimensional feature flow. If it is higher than 5ms, the implicit phase delay accumulated by the underlying hardware while waiting for communication alignment will exceed the upper limit of the pipeline state machine's buffer queue capacity, causing the system to crash.

[0040] Example 3: In a general-purpose computing platform operating in a computing cluster environment with a floating-point computing power of 100 trillion floating-point operations per second or higher, this example is used to handle a situation where the nested loops of control flow logic reach 8 layers and the external hardware communication path experiences a sudden phase delay of 25ms. At this time, unidirectional transmission resistance occurs between cascaded operator layers, causing the computational features of neurons in the feature space to diffuse and inducing topological divergence in the phase space manifold of the computational model. The operator pipeline triggers a cascading amplification of temporal mismatches in the data transmission chain, generating computational bubbles within the neural network computation graph and causing the computation process to suspend and temporal logic deadlock failures. The data acquisition unit collects the branch prediction probabilities of multi-level action nodes in the heterogeneous software system and the directed topological flow data of the business logic as the baseline input. The multi-level neural network processing module converts the directed topological flow data of business logic into a high-dimensional feature tensor containing a logical dependency pointer. The computing unit measures the statistical variance of the high-dimensional feature tensor within a preset sliding time window containing 128 consecutive sampling points. To quantify the deviation of the feature manifold caused by temporal asynchrony during the inference process of each level of operator, the computing unit divides the feature space distribution residual by the statistical variance within the time sliding window, and the ratio is defined as the temporal correlation dispersion parameter. When the statistical variance of the first feature channel in the data processing node is 0.74 and the measured feature space distribution residual is 0.65, the computing unit divides the feature space distribution residual of 0.65 by the statistical variance of 0.74, and calculates the value of the temporal correlation dispersion parameter as 0.878.

[0041] When the time-series correlation dispersion parameter is greater than or equal to the set safety threshold of 0.85, the control unit closes the external weight input channel to ensure that the update step size of the external weight input is always equal to 0. Simultaneously, it triggers the data feedforward decoupling and reconstruction process. The central processing unit extracts the state prediction probability tensor of the current step test action and subtracts the constant-coefficient steady-state reference tensor stored in the historical data section to generate a prediction deviation matrix. The phase delay difference caused by the 25ms physical delay jitter measured by the system timer is added to a constant 1 to generate a step denominator. The prediction deviation matrix is ​​divided by the step denominator to calculate the in-situ adjustment increment in the corresponding weight adjustment matrix. The front-end neural network processing module calculates the nonlinear Hadamard product based on the in-situ adjustment increment and the high-dimensional feature tensor to generate the core feature matrix, thereby completing the dynamic allocation of weights within the front-end neural network processing module. The in-situ overlay modification of the matrix and the orthogonal projection transformation of the topological eigenvalues ​​of the output core feature matrix are performed according to preset boundary constraints. In practice, the specific steps of the orthogonal projection transformation include: the central processing unit performs singular value decomposition on the core feature matrix to extract eigenvalue vectors that represent the topological expansion features of the computation graph; then, the eigenvalues ​​are hard-truncated using preset boundary constraints, which are set so that the absolute values ​​of the eigenvalues ​​are within a closed interval of 0 to 1; if an eigenvalue exceeds 1, it is forcibly projected and scaled to 1, and if it is less than 0, it is fixed to 0, thereby constructing a unit orthogonal projection transformation matrix; finally, the adjusted eigenvalue vectors are multiplied and recombined with the original eigenvector matrix to restore the new feature matrix, thereby forcibly constraining the divergent manifold of the eigenvalues ​​within a closed convex polytopic topological space defined by a multivariate orthogonal unit basis.

[0042] In the interlayer data flow loop between the in-situ weight update of the front-end neural network processing module and the intermediate pipeline state machine, boundary constraints confine the topologically divergent computation graph to a preset closed topological space. The output processing module receives the core feature matrix output after orthogonal projection transformation and calls the script syntax control operator to process the data flow topology rearrangement. The script syntax control operator adjusts the spatial arrangement of control flow nodes and aligns the temporal sequences in the time-dependent graph. Simultaneously, the output processing module calculates the linear evaluation index of the priority based on the instruction type and control depth of the script syntax control operator, and uses the linear evaluation index to prioritize each control instruction. The gender mapping rule directly superimposes the resulting convergence point weight adjustment tensor onto each control flow node in the timing dependency graph, outputting test script control command flow to the automated test equipment to drive the heterogeneous software system to generate automated test actions. The forward hierarchical inference latency of the neural network computation graph is maintained at 4.65ms. The timing mismatch caused by the unidirectional cascading of multi-level operators is eliminated. The computing system operates in a bus characteristic bandwidth environment with 500 to 2000 calls per second and its characteristic matrix data throughput is stable at 850B / s. The state convergence time of the computation pipeline when the number of nested loops of the control flow logic conditions reaches 8 layers is reduced from the divergent state to 16.8ms.

[0043] Example 4: When the system faces the initial deployment scenario of migrating cross-platform automated testing services to a new testing environment, the dedicated computing platform needs to handle the flow discontinuity caused by the undefined initial baseline of the state prediction probability tensor in the front-end neural network processing module before the directed topology data is entered; the processor adjusts the test operator cluster to an idle state, and within a continuous 5-minute sampling period, the data acquisition unit obtains the system latency of the test host at a sampling frequency of 200Hz. The system computing unit reads the dynamically allocated weight matrix of the initial state, and by statistically analyzing the cumulative deviation of the feedforward calculation of each action layer, it performs amplitude normalization and alignment of the mapping weights of neurons at each level, completing the back-end alignment within the computation graph. The ablation mechanism of background noise features is as follows: When the test host is in an idle state, the system latency jitter is mainly caused by the task scheduling and hardware interrupts of the underlying operating system. In the feedforward calculation, it manifests as random and disordered amplification of the feature amplitudes of each channel of the weight matrix. By normalizing the magnitude of the row vectors of the mapping weight matrix of each action layer, the amplitude distortion introduced by the asynchronous disturbance of the hardware environment is forcibly scaled to a uniform unit scale benchmark, so that the feedforward response depends only on the relative structural changes of the business topology flow. Thus, the transmission components of low-amplitude high-frequency latency noise are mathematically filtered out and canceled within the neural network, achieving in-situ ablation of background noise features.

[0044] With the completion of idle baseline alignment, the computing unit calculates the mean of the sampled steady-state phase space neuron state distribution tensor and saves the calculated steady-state probability matrix as a constant-coefficient steady-state reference tensor to the physical reference address of the non-volatile memory. The idle delay baseline reading recorded by the system timer is in the range of 0.12ms to 0.15ms. The constant-coefficient steady-state reference tensor and the idle delay baseline reading are used as inputs for calculating the safety critical threshold, so that the intermediate pipeline state machine has a static reference for deviation comparison at the first time it receives data, and the interlayer propagation resistance of the computing operator in forward inference is maintained below the set safety value.

[0045] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for automatic script generation based on a pipelined neural network architecture, characterized in that, Includes the following steps: Step S1: Obtain the business logic directed topology flow data and the branch prediction probability of multi-level action nodes. Use the front-end neural network to map the business logic directed topology flow data into a high-dimensional feature tensor containing a logical dependency pointer. Extract the variance features of the high-dimensional feature tensor within the sliding time window. Multiply the branch prediction probability of the multi-level action nodes with the logical dependency pointer and calculate the normalized difference comparison value. Output the dimensionless temporal correlation dispersion parameter. Step S2: In the hierarchical feedforward calculation of the multi-level neural network, the data transmission delay status between layers is determined. When the dimensionless temporal correlation discreteness parameter is greater than the safety critical threshold of 0.85, the data feedforward decoupling and reconstruction process is triggered. The external weight input channel is disconnected, and the dynamic weight allocation matrix in the previous neural network is covered and modified in situ using the weight adjustment matrix. The in situ adjustment increment matrix in the weight adjustment matrix is ​​calculated and nonlinear Hadamard product is obtained with the high-dimensional feature tensor. The topological structure of the computation graph is constrained to the topological space. Step S3: Receive the reconstructed high-dimensional feature tensor, configure the script syntax control operator to complete the data flow topology rearrangement, determine the convergence point weight adjustment amount using matrix linear mapping rules when determining the priority of control instruction generation, and finally generate test script code.

2. The script automatic generation method based on a pipelined neural network architecture according to claim 1, characterized in that, In step S2, determining the inter-layer data transmission delay status in the hierarchical feedforward calculation of the multi-level neural network includes the following steps: Step S21, when the hierarchical inference delay is greater than the critical delay threshold of 5ms, the dimensionality reduction hedging process is started. Step S22: Quantize the time-series phase difference into a delay compensation weight increment matrix, call the constant coefficient steady-state reference tensor saved in the historical data section to dynamically allocate the weight matrix, use the static topological constraint matrix to complete the feature slicing of the high-dimensional feature tensor, and use the discretized gradient to eliminate the inter-layer cumulative delay.

3. The script automatic generation method based on a pipelined neural network architecture according to claim 1, characterized in that, Step S1 includes the following sub-steps: Step S11, obtain the branch prediction probability of the multi-level action node; Step S12, perform normalized difference comparison on the product of the branch prediction probability of the multi-level action node and the logical dependency index, and output the dimensionless temporal correlation dispersion parameter.

4. The script automatic generation method based on a pipelined neural network architecture according to claim 1, characterized in that, Step S2, which puts the external weight input into a frozen state, includes the following sub-steps: Step S23, after receiving the in-situ adjustment increment matrix, disconnect the channel of the external weight input; Step S24, lock the update step size of the external weight input to 0, and control the weight update of the previous neural network to respond independently to the in-situ adjustment increment matrix.

5. The script automatic generation method based on a pipelined neural network architecture according to claim 1, characterized in that, Step S2, which involves calculating the nonlinear Hadamard product based on the in-situ adjustment increment matrix and the high-dimensional feature tensor and constraining the computation graph topology to the topological space, includes the following sub-steps: Step S25, calculate the element-wise product of the in-situ adjustment increment matrix and the high-dimensional feature tensor, and output the decoupled core feature matrix; Step S26, perform orthogonal projection on the eigenvalues ​​of the core feature matrix using boundary constraints, and constrain the computation graph topology to the topological space.

6. The script automatic generation method based on a pipelined neural network architecture according to claim 2, characterized in that, Step S22 involves slicing the high-dimensional feature tensor using a static topological constraint matrix and eliminating inter-layer cumulative delay using discretized gradients. This includes the following sub-steps: Step S221: Use a static topological constraint matrix to intercept the topological expansion of the high-dimensional feature tensor during forward propagation; Step S222: Slice the high-dimensional feature tensor along the time axis, calculate the differential gradient between adjacent slices, and eliminate inter-layer data transmission delay using the differential gradient.

7. The script automatic generation method based on a pipelined neural network architecture according to claim 1, characterized in that, The configuration of script syntax control operators to complete the data flow topology reordering in step S3 includes the following sub-steps: Step S31, construct the temporal dependency graph of script syntax based on the reconstructed high-dimensional feature tensor; Step S32, use script syntax control operators to adjust the spatial arrangement and time axis alignment of the control flow nodes in the temporal dependency graph.

8. The script automatic generation method based on a pipelined neural network architecture according to claim 7, characterized in that, Step S3, in determining the priority of control command generation, uses the matrix linear mapping rule to determine the convergence point weight adjustment amount, including the following sub-steps: Step S33, extract the command type and control depth of the script syntax control operator, and construct a linear evaluation index for priority; Step S34, use the linear evaluation index to sort the priority of each control command, and calculate the convergence point weight adjustment amount in the time-series dependency graph.

9. The script automatic generation method based on a pipelined neural network architecture according to claim 8, characterized in that, The generation of test script code includes the following sub-steps: Step S35, apply the convergence point weight adjustment to each control flow node in the time-series dependency graph; Step S36, output the test script control instruction flow according to each control flow node after spatial arrangement adjustment and time axis alignment.

10. The script automatic generation method based on a pipelined neural network architecture according to claim 2, characterized in that, The initialization, deployment, and storage steps for the constant-coefficient steady-state reference tensor stored in the historical data section include: under the initialization deployment condition where no directed topology data has been entered, adjusting the test operator cluster to an idle state, obtaining the system latency of the test host, and reading the dynamic weight allocation matrix of the initial state; by statistically analyzing the cumulative deviation of the feedforward calculation of each action layer, normalizing and aligning the row vectors of the mapping weights of neurons at each level, and absolving background noise features in situ; after the idle baseline alignment, calculating the mean of the sampled steady-state phase space neuron state distribution tensor, and using the calculated steady-state probability matrix as the constant-coefficient steady-state reference tensor, which is then stored in the physical reference address of the non-volatile memory.

Images