Welding robot on-chain credible multi-expert gating self-optimization control method and system

Abstract

The invention discloses a credible multi-expert gating self-optimization control method and system on a welding robot chain. The system is composed of an edge side and a cloud side, wherein the edge side comprises data acquisition and feature extraction, multi-expert control, sparse gating routing, compliance agent, security agent, execution and drive, block chain client and degradation control; the cloud side provides strategy service, a rule base and an audit interface. The method comprises the steps of collecting multichannel signals to obtain working conditions and confidence, generating control increments by multiple experts, performing gating sparse selection / weighting, performing compliance cutting and security check'double signatures' issuing, generating'parameter-quality 'abstract uplink and triggering events and settlement in case of abnormity, and performing local PID + amplitude limiting and chain supplement recovery and smooth switching in case of strategy disconnection. According to the method, preferential control according to working conditions, rigid compliance and traceable treatment are achieved, splashing and out-of-limit risks are reduced, and energy efficiency and continuity are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and welding automation, specifically to a trusted multi-expert gating self-optimization control method and system for welding robot chains, belonging to the cross-technology of industrial process control, edge-cloud collaboration and trusted computing. Background Technology

[0002] The existing welding robot production line has the following pain points: (1) Process drift and parameter overrun: Welding current / voltage, wire feeding and walking speed are prone to slow drift or sudden changes in long welds and multi-layer multi-pass conditions, resulting in increased spatter, insufficient penetration or abnormal heat-affected zone. (2) Black box optimization and lack of traceability: Parameter tuning based on experience or black box models is difficult to review, and the responsibility for quality disputes is poorly defined and reproducible. (3) Inconsistency of multi-source sensors: Multi-source signals such as electrical parameters, infrared thermography, arc light spectrum and acoustic emission have time delay / mismatch / noise, and direct fusion is prone to introducing artifacts. (4) Cloud-side strategy unavailability: When network jitter and strategy services are unavailable, online optimization is interrupted, and security degradation is required to ensure continuous production. (5) Ambiguous equipment sharing and compliance billing: There is a lack of settlement and compliance proof based on data facts in the equipment sharing scenario across enterprises / workshops.

[0003] Therefore, an integrated solution is needed that simultaneously possesses **(i) multi-expert control and sparse gating routing, (ii) compliance and security dual-signature constraints, (iii) on-chain traceability and event-based settlement, and (iv) cloud-edge collaboration and security degradation**. Summary of the Invention

[0004] 1. Purpose of the invention This invention aims to address the pain points in welding robot production lines, such as uncontrollable parameter exceeding limits and process drift, untraceable black-box optimization, inconsistent multi-source sensing, interruption caused by cloud-side strategy unavailability, and difficulties in settlement and compliance proof in equipment sharing scenarios. It proposes a blockchain-based trusted multi-expert gating self-optimization control method and system. Under the premise of ensuring process safety and regulatory compliance, it achieves the following: (1) Selecting the most suitable expert control strategy according to the working conditions to suppress spatter, stabilize the molten pool, constrain heat input, and reduce energy consumption; (2) Implementing rigid constraints and fallback measures for control commands with dual signatures of compliance agent + security agent; (3) Anchoring the "parameter-quality" summary of the entire process with blockchain, and conducting event-based governance and linkage settlement when anomalies occur; (4) Smoothly switching between security degradation and recovery when cloud-side strategy is disconnected, ensuring continuous production and auditability and reproducibility.

[0005] 2 Overall Technical Solution This invention proposes a cloud-edge collaborative architecture: The edge deployment includes modules for data acquisition and feature extraction, multi-expert control, sparse gating routing, compliance and security dual agents, execution and driving, blockchain client, and degradation control. The cloud side provides policy services (expert / gating training and issuance), rule base version management, audit interfaces, and event / settlement smart contracts. Within each control cycle, the system completes: data collection → feature / confidence calculation → expert advice → gating sparse routing → compliance pruning → security verification and dual signature → execution → on-chain data entry and events, and initiates degradation and on-chain replenishment in case of anomalies.

[0006] 3 System Technical Solution The system of the present invention includes the following modules and their data / control relationships: (A) Data acquisition and feature extraction module: synchronously acquires current, voltage, wire feeding speed, travel speed and at least one molten pool related signal (infrared thermography, arc spectroscopy, acoustic emission, etc.) at a frequency of not less than 200 Hz (preferably 500–2000 Hz), and outputs the working condition vector x in a sliding window of 100–500 ms. t With sensor confidence g t (Estimated from SNR, consistency, and missing rate). (B) Multi-expert control module: Includes at least four types of experts: melt pool stabilization, spatter suppression, heat input constraint, and energy consumption optimization (optional equipment health expert), for x t Output candidate control increment yᵢ, with the interface uniformly set to {channel ID, increment value, confidence level}. (C) Sparse gated routing module: Input concat(x t , g t(A) Obtain the sparsity coefficient α, constrain **||α||0≤2**, and set the hysteresis threshold and minimum dwell time to suppress frequent expert switching jitter; output the synthetic increment Δu*. (D) Compliance agent module: Based on the rule base R, hard constraint pruning is applied to Δu*, including upper and lower bounds, maximum rate of change, unit consistency, heat input HI interval, posture / level / bevel boundary, energy budget / carbon emission estimation, etc., to obtain Δu′; when there are multiple constraint conflicts, the pruning result is generated according to priority or multi-objective trade-off. (E) Security agent module: Based on the historical trajectory buffer H, the timing consistency, jitter, and actuator saturation check are performed on Δu′; if necessary, amplitude limiting / filtering / freezing and rollback are applied. The compliance agent and security agent sign Δu′ respectively, and execution can only be issued after forming a double signature. (F) Blockchain Client Module: Generates "parameter-quality" segmented summaries for each weld pass / pass and aggregates and anchors them on the chain using a Merkle tree; triggers on-chain events for anomalies (exceeding limits, rework, sensor mismatch, communication anomalies); in equipment sharing scenarios, it links with the settlement smart contract for metering and billing. In consortium blockchain scenarios, details are on-chain and hash is on-chain, achieving privacy isolation and non-repudiation. (G) Degradation Control Module: When the cloud-side policy is unavailable, it switches to the local PID + limit + maximum slope policy; after the cloud side recovers, it smoothly switches back to gating control using time-weighted fusion and replenishes the chain for the offline summary. (H) Cloud-side Policy Service: Responsible for expert / gating training and distillation compression, model issuance and verification, rule base R version management and audit interface.

[0007] 4. Methods and Technical Solutions A reliable self-optimizing control method for a welding robot chain includes: S1 Acquisition and Feature Analysis: Synchronously acquiring multi-channel signals at a frequency of not less than 200Hz; extracting x using a sliding window of 100–500 ms. t And calculate g t S2 expert recommendation: Multiple experts should conduct parallel analysis on x. t Generate yᵢ. S3 Sparse Gating: The gating network is based on concat(x) t , g t S4. Compliance Pruning: Based on the rule base R, prune Δu* according to the upper and lower bounds, maximum rate of change, **HI=η·V·I / TravelSpeed ​​(kJ / mm)** interval and unit consistency to obtain Δu′. S5. Security Verification and Dual Signature: Perform time-series consistency, jitter, and saturation verification on Δu′; after processing, authorize with dual signatures with the compliance agent and issue for execution. S6. On-Chain and Event / Settlement: Generate segmented digests and aggregate them on-chain using a Merkle tree; trigger events in case of anomalies and link them with the settlement contract (e.g., fee adjustment / pause). S7. Degradation and Smooth Switching: Execute PID+limiting when disconnected; after recovery, smoothly switch back to gating within the T_blend window and replenish the offline digest on-chain.

[0008] 5. Key technical features and parameter ranges (supporting implementation feasibility and repeatability) T1 Sampling and Window: f s ≥200 Hz (preferably 500–2000 Hz); window W=100–500 ms, step size Δ=50–100 ms. T2 Sparse gating: ||α||0≤2; typical hysteresis thresholds θ_on≈0.6, θ_off≈0.4; minimum dwell time τ_min≈0.5–2 s. T3 Maximum rate of change: r_max is configured for both travel speed and wire feeding speed (example: travel ≤4 mm / s², wire feeding ≤0.4 m / min·s⁻¹, can be set according to equipment rating). T4 HI range: set according to material / plate thickness / posture, such as 8 mm butt joint PA for low carbon steel: HI∈[0.8,1.6] kJ / mm (η∈[0.6,0.9]); unit is uniformly kJ / mm. T5 Safety Verification Criteria: Chattering is determined by the frequency band energy threshold; saturation is determined by the actuator duty cycle / current margin; the handling is amplitude limiting or low-pass filtering. T6 On-Chain Anchoring: Segmented summaries of the same weld pass are uploaded to the chain using Merkle roots; event types include at least ParamBreach, SensorMismatch, ReworkIssued, NetAnomaly, and LeaseBilling; the consortium blockchain supports off-chain encrypted storage of details and on-chain hash anchoring. T7 Degradation Strategy: PID uses ZN initialization and superimposed amplitude limiting and maximum slope; smooth switching uses time-weighted fusion (T_blend typically 1–5 s).

[0009] Evaluation of beneficial effects and technical effects Compared with the prior art, the present invention achieves the following effects: E1 optimal control based on operating conditions: through g t Participation in gating suppresses the impact of noisy or mismatched sensor channels, sparsely activating 1-2 of the most relevant experts and reducing "general compromises." E2 Rigid Compliance and Security Backup: Hard constraints such as unit / range / maximum rate of change / HI interval and dual signatures for security verification significantly reduce the risk of exceeding limits and actuator impact. E3 Traceability and Event-Based Governance: Segmented summaries + Merkel aggregation + on-chain events enable auditable quality disputes, measurable and non-repudiable equipment sharing. E4 Continuity and Resilience: Automatic degradation upon cloud-side disconnection, smooth switching and chain replenishment upon recovery, reducing downtime and rework. E5 Improved Metrics (Example): Under typical GMAW docking conditions, splash count is reduced, energy consumption / meter is reduced, and the delay in handling exceeding limits remains at the millisecond level; specific values ​​depend on the operating conditions and equipment.

[0010] Comparison with existing technologies and key creative points C1. “Gated MoE + Sensor Confidence” Online Routing Mechanism: Unlike single-expert or fixed-weight fusion, it suppresses frequent switching and noise amplification through sparse gating, hysteresis, and dwell. C2. “Compliance Agent + Security Agent” Dual-Signature Rigid Constraints: Unlike optimization control relying solely on soft objectives, it authorizes execution under multi-dimensional hard constraints such as unit consistency, HI interval, maximum rate of change, and pose / layer boundary. C3. “Segmented Summary + Merkel Aggregation + Event-Based Settlement” On-Chain Trusted Closed Loop: Unlike simple notarization, it links process anomalies with settlement contracts to form a governance closed loop. C4. “Cloud-Edge Collaboration + Degradation and Smooth Switching” Resilience Control: Unlike single-end strategies, it ensures that the strategy does not degenerate into an unconstrained state when it is unavailable and can be recovered in an auditable manner. C5. Strongly Typed Rule Base and Unit / Range Guardrails: Provide machine-readable and verifiable constraints on key welding process quantities (HI, interlayer temperature, energy budget), reducing “black box” uncertainty. Attached Figure Description

[0012] Figure 1 System overall architecture diagram: showing the logical connection, data and control paths between the cloud, edge and robot controller and sensors.

[0013] Figure 2 Data flow and state machine: Describe the sequential relationship and state transitions of data collection → feature generation → multi-expert access → gating → compliance → security → execution → on-chain.

[0014] Figure 3 Sparse gated routing diagram: illustrating the generation of weight vector α, expert selection / weighting, and hysteresis control mechanism.

[0015] Figure 4 Dual-signature process for compliance agent and security agent: describes the sequence of unit / scope / HI / change rate verification, security check and dual-signature authorization.

[0016] Figure 5 On-chain event and settlement sequence diagram: illustrating over-limit / return / sensor mismatch / communication anomaly → event → contract settlement (including chain replenishment and channel isolation).

[0017] Figure 6 Degraded control and smooth switching process: Example PID + limiting → time-weighted fusion → gating switching path.

[0018] Figure 7 Example timing curves: showing the changes in current / voltage / speed / thermal input (HI) over time and event markers. Detailed Implementation

[0019] To make the technical solution of the present invention clearer, the "Reliable Multi-Expert Gated Self-Optimizing Control Method and System for Welding Robot Chain" will be described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the described embodiments are for illustrative purposes only and not for limiting the invention; various modifications and substitutions can be made by those skilled in the art without departing from the spirit of the invention.

[0020] I. System Architecture and Hardware / Software Environment 1. Overall architecture of cloud-edge collaboration (see Figure 1 ) This invention adopts a layered cloud-edge collaborative structure. The edge side comprises industrial computing units located near the welding robot control cabinet, connected to the robot controller, welding power supply, and multi-source sensors (electrical parameters, infrared thermography, arc light spectroscopy, acoustic emission / vibration, etc.) via industrial Ethernet or fieldbus. The cloud side deploys services such as model training and issuance, rule base version management, auditing, and consortium blockchain ledger node interfaces. The network uses gigabit Ethernet or a 5G private network, with a desired one-way latency of <20 ms; in the event of a connection failure, the edge side possesses data buffering and connection replenishment capabilities.

[0021] 2. Hardware and software stack and time synchronization On the edge, Linux (with optional real-time kernel) can run, deploying a data acquisition SDK, inference engine (such as ONNXRuntime / TensorRT), blockchain client SDK, and containerized services. On the cloud, deep learning training frameworks (such as PyTorch / TF), model repositories, signature / verification and rule base management services, and consortium blockchain nodes run. PTP / IEEE 1588 master-slave clock synchronization is used to ensure consistency between sampling and on-chain timestamps.

[0022] II. Functional Modules and Data Flow (see...) Figure 2 ) 1. Data Acquisition and Feature Extraction Module (Module 1) Acquisition channels: current I, voltage V, wire feed speed v_w, travel speed v_t; optional thermal imaging ROI statistics, spectral characteristic line intensity ratio, acoustic emission short-time energy, etc.

[0023] Sampling and alignment: For electrical and motion-related channels, the preferred Hz is 500–2000 Hz (minimum 200 Hz); thermal images and spectra are acquired at frame rates and time-aligned to the control cycle.

[0024] Sliding window and features: Time-domain statistics (mean, variance, skewness, kurtosis), frequency-domain energy ratio, texture and gradient, spectral line ratio, envelope peak spacing, etc. are calculated with a window width of 100–500 ms and a step size of 50–100 ms to form a feature vector; a confidence score of 0–1 is obtained based on SNR, channel consistency and missing rate, which is used for subsequent gating weighting.

[0025] 2. Multi-expert control module (Module 2) An expert ensemble is constructed, including but not limited to: pool stabilization (E_pool), spatter suppression (E_spat), heat input constraint (E_hi), energy consumption optimization (E_energy), and optional device health (E_health). Each expert outputs incremental suggestions and confidence levels for control channels (wire feeding, travel, oscillation amplitude / frequency, etc.) through a unified interface. The expert model is a lightweight network, with edge-side inference and periodic distillation and issuance on the cloud side.

[0026] 3. Sparse gated routing module (Module 3, see [link]) Figure 3 ) The gating network takes "feature + confidence" as input and outputs the weights of each expert. The system enforces sparse constraints in each control cycle (activating a maximum of 1-2 experts per cycle) and sets start-stop thresholds and minimum dwell time to suppress jitter. The suggestions of the activated experts are linearly synthesized according to their weights to obtain the comprehensive control increment.

[0027] 4. Compliance Agency Module (Module 4, see...) Figure 4 ) Based on the rule base R, the comprehensive increment is checked for: unit consistency, upper and lower bound pruning, maximum rate of change limit, heat input HI interval check (HI=η·V·I / TravelSpeed,kJ / mm,η∈[0.6,0.9]), and matching check of posture / level / bevel / energy budget and carbon emission constraints; if there is a conflict, adjustments are made in accordance with the principle of "legal standards > enterprise specifications > historical experience" and "minimum change principle".

[0028] 5. Security Proxy Module (Module 5, see [link]) Figure 4 ) Maintain a historical buffer to assess timing consistency, chattering (high-frequency energy threshold), and actuator saturation; for anomalies, recommend limiting, low-pass filtering, or freeze rollback. The compliance agent and security agent each digitally sign the final control quantity, allowing distribution only when both signatures are valid.

[0029] 6. Execution and Driver Module (Module 6) The control increments of dual-signature authorization are mapped to the controller registers / interfaces; hardware limiters and maximum slopers are configured on the controller side to form the final physical security boundary.

[0030] 7. On-chain Trust and Event / Settlement Module (Modules 7 / 8, see also) Figure 5 ) A summary is generated for the effective control segments and key quantities, and anchored on the consortium blockchain in the form of a Merkle tree root; details are stored in encrypted off-chain storage. The system identifies and reports events such as ParamBreach, SensorMismatch, ReworkIssued, NetAnomaly, and LeaseBilling, driving the contract to perform billing or rate adjustment / suspension.

[0031] 8. Degradation and smooth handover module (Module 9, see...) Figure 6 ) When the cloud side is unreachable, signature verification fails, or network quality is below the threshold, it automatically degrades to PID + amplitude limiting + maximum slope control. After the cloud side recovers, it smoothly transitions from the degraded output back to the gated output with time weighting within the set fusion window, while summarizing the offline period and recording abnormal events.

[0032] 9. Cloud-side policy service (Module 10) Responsible for model training, distillation and quantization, signature / verification, rule base version management and auditing; only distribute model packages that have passed signature verification.

[0033] III. Step-by-step description of key processes (without pseudocode) 1. Feature and confidence score calculation (corresponding) Figure 2 (data stream) (1) Extract the most recent time window from the acquisition buffer; (2) Calculate the time domain and frequency domain features of the electrical and motion channels; (3) Calculate the statistics and gradient of the thermal image ROI, the characteristic spectral line ratio of the spectrum, and the acoustic emission short window features; (4) Concatenate them into a feature vector; (5) Obtain a 0–1 confidence level based on SNR, inter-channel consistency, and missing rate, and use it as the reliability weight of the gated input.

[0034] 2. Sparse gating and expert routing (see...) Figure 3 ) (1) Output expert weights from the gating network; (2) Perform sparsification and hysteresis / residential control, limiting the number of activated experts per cycle to ≤2; (3) Collect incremental suggestions from activated experts for each channel; (4) Synthesize the overall control increment linearly according to the weights.

[0035] 3. Compliance and Security: Dual Proxy and Dual Signature (see...) Figure 4 ) (1) The compliance agent completes the verification of the unit, upper and lower limits, maximum change rate and HI interval, and performs constraint mapping on the pose / bevel / layer matching; (2) When conflicts occur, the agent is trimmed according to the priority and minimum change principle; (3) The security agent determines the chattering, saturation and timing consistency based on the historical buffer, and implements amplitude limiting / filtering / freezing for anomalies; (4) The compliance agent and the security agent sign the final control quantity respectively, and the issuance is only allowed when the double signature is valid.

[0036] 4. On-chain evidence storage, event identification, and settlement (see...) Figure 5 ) (1) Calculate the summary of the control segment and key quantity and update the Merkle root, and anchor it on the chain; (2) Trigger event objects such as over-limit, rework, sensor mismatch and communication abnormality; (3) Link with the billing contract to automatically settle or adjust the rate based on duration, energy consumption and service level; (4) Only store the evidence summary and necessary metadata on the chain to protect process privacy.

[0037] 5. Degradation control and smooth handover (see...) Figure 6 ) (1) When the cloud side is unavailable, immediately switch to the PID+limiting+maximum slope safety strategy; (2) After the cloud side recovers, set a fusion window of 1–5 s and transition from degraded output to gated output according to time weight; (3) Supplement the offline data and label it with NetAnomaly.

[0038] IV. Parameter Range and Engineering Settings Sampling rate: 500–2000 Hz (minimum 200 Hz) is preferred for electrical activity-related subjects. Sliding window / step size: W=200 ms (adjustable 100–500 ms), Δ=50–100 ms; Sparse gating: The number of experts activated per cycle is k=1–2, with suggested values ​​of θ_on=0.6, θ_off=0.4, and minimum dwell time τ_min=0.5–2 s; Example of maximum rate of change: walking ≤ 4 mm / s², wire feeding ≤ 0.4 m / min·s⁻¹; Heat input range: determined by the rule library according to material / thickness / position; example (low carbon steel 8 mm, PA position) is 0.8–1.6 kJ / mm, η=0.8; Smooth transition: Blend window T_blend = 1–5 s.

[0039] The above parameters can be incorporated into the relevant clauses of the claims as dependent scopes or preferred values.

[0040] V. Blockchain Anchoring and Privacy Protection (see...) Figure 5 ) Evidence granularity: A summary is generated for each weld bead segment, and after Merkel aggregation, it is uploaded to the blockchain to ensure traceability; Privacy policy: Different production lines / customers use independent channels for isolation; only summaries and necessary metadata are stored on-chain, while details are encrypted and stored off-chain, with access traces and audits. Event-driven events such as ParamBreach, SensorMismatch, ReworkIssued, NetAnomaly, and LeaseBilling can trigger rate adjustments, suspension of settlements, or rework processes.

[0041] Example 1. Operating conditions and equipment The material is 8 mm thick low-carbon steel plate with a butt weld of approximately 1.2 m; GMAW process, wire feed diameter 1.2 mm, mixed gas protection. Electrical participation motion is sampled at 1 kHz; thermal imaging is taken at 50 Hz using the mean / variance / gradient of the ROI; the spectrum is taken as the ratio of two characteristic spectral lines; acoustic emission is sampled at 20 kHz and characterized to 1 kHz alignment.

[0042] 2. Rule base (in this example) Upper and lower limits: current 120–260 A, voltage 20–32 V, travel 3–10 mm / s; maximum rate of change: travel ≤4 mm / s², wire feed ≤0.4 m / min·s⁻¹; HI range 0.8–1.6 kJ / mm (η=0.8); posture PA, interlayer temperature ≤180℃.

[0043] 3. Operation process Start-up (0–20 s): Establish the molten pool and thermal equilibrium using degradation control, and record the "start-up" event; Gating (20–45 s): Gating prioritizes the activation of melt pool stability and thermal input experts (typical weights approximately {0.7, 0.3}). As thermal image noise increases, the confidence level decreases, causing gating to automatically reduce the weight of this information source. Disturbance and Handling (≈45 s): When splash increases suddenly, splash suppression experts recommend reducing the current and moderately increasing the walking speed; the compliance agent limits the amplitude according to the maximum rate of change, and the security agent issues the signal after confirming no jitter / saturation and double-signing with the compliance agent; ParamBreach is recorded on the chain (warning → handling); Network interruption (60–62 s): switch to degraded control; after recovery, smoothly switch back to gating within a 3 s fusion window and record NetAnomaly and complement summary; Final step: After Merkle aggregation of the fragment summary, it is uploaded to the blockchain, and the contract generates a LeaseBilling record.

[0044] 4. Exemplary Effects Compared to an ungated baseline, this embodiment shows a splash count reduction of approximately 15–20%, an out-of-limit handling delay of <120ms, and a power consumption reduction of approximately 5–8% per unit length. The above data are illustrative; actual results are dependent on equipment, materials, and orientation.

[0045] Alternative implementation methods and variations The expert pool can be expanded to form quality / deformation control, or simplified to {E_pool,E_spat} in the high-speed welding scenario of thin plates. The gating implementation can be replaced with a lightweight attention or gated MoE, but keep the number of activated experts per cycle ≤2; The rule base can manage material-welding material-position-thickness combination constraints in a graphical manner, and incorporate energy consumption / carbon emission budgets; The ledger backend can be replaced with different consortium blockchains / sidechains, without affecting the key points of "segmented summary - Merkel anchoring - event linkage"; The degradation strategy can employ a safe subset of MPC with maximum slope and amplitude constraints.

[0046] Engineering Considerations and Best Practices Time synchronization: PTP takes precedence; if the time out of sync is greater than 10 ms, correction is triggered and noted in the abstract; Unit / Scope Guardrail: Strong type checking is implemented at the edge to reject data with undefined units from being entered into the database or uploaded to the blockchain; Model security: If the policy package fails to verify the signature, it will not be loaded and will revert to the verified version. Privacy compliance: Only certificate summaries and necessary metadata are stored on-chain, while off-chain details are encrypted and access is audited; Validation of parameters: It is recommended to first calibrate the maximum rate of change, HI range and gating hysteresis threshold in virtual welding or small samples before putting it on the production line.

[0047] The present invention has been described in detail above; any equivalent substitutions and modifications made to the structural layout, algorithm details, parameter ranges and on-chain implementation without departing from the core idea of ​​the present invention shall fall within the protection scope of the present invention.

Claims

1. A reliable multi-expert gated self-optimizing control method for a welding robot chain, characterized in that, Includes the following steps: a) Synchronously acquire current, voltage, wire feed speed, travel speed, and at least one melt pool-related signal at a frequency of not less than 200 Hz. Calculate the operating condition feature vector based on a sliding time window of 100 to 500 milliseconds and a step size of 50 to 100 milliseconds, and obtain a confidence level between zero and one based on signal-to-noise ratio, channel consistency, and missing rate; b) Generate candidate incremental suggestions for control channels such as wire feed, travel, and oscillation in parallel by at least two expert models, and provide their respective confidence levels; c) Route the features and confidence levels through a gating network, implement sparsity constraints to ensure that the number of experts activated in each control cycle does not exceed two, and set start-stop-back thresholds and minimum dwell time to suppress frequent switching. Linearly synthesize the suggestions of the activated experts according to their weights into a comprehensive control increment; d) Input the integrated control increment into the compliance agent to complete unit unification, upper and lower bound clipping, maximum rate of change limitation, and thermal input constraint verification. Thermal input equals η multiplied by voltage multiplied by current divided by travel speed, in kilojoules per millimeter (kJ / mm). η ranges from 0.6 to 0.

9. Simultaneously, matching and verification are performed based on posture, level, slope, energy budget, and carbon emission constraints to obtain the compliant control increment. e) Input the compliant control increment into the safety agent. Based on historical trajectory buffers, it judges timing consistency, chattering, and actuator saturation, and implements amplitude limiting, filtering, or freeze rollback as necessary. Both the compliance agent and the safety agent digitally sign the final control increment. f) Issue the control fragments and key quantities that have been issued and are effective to the controller. Use Merkle tree aggregation to update the root digest of the corresponding weld and anchor it on the consortium blockchain. At the same time, generate events for states such as exceeding limits, rework, sensor mismatch and communication anomalies and link them with the settlement contract for billing or rate adjustment. g) When the cloud-side strategy is unreachable or the signature verification fails or the network quality is lower than the threshold, automatically switch to local proportional-integral-derivative control and superimpose amplitude and maximum slope constraints. After the cloud side recovers, smoothly transition from degraded control to gated output in a time-weighted manner within a time window of one to five seconds, and supplement the digest generated during the offline period on the blockchain.

2. The method according to claim 1, characterized in that, Sampling and feature extraction include: calculating the mean, variance, skewness, kurtosis, and band energy ratio for electrical and motion channels; calculating the mean, variance, and gradient for a specified region of infrared thermography; calculating the characteristic spectral line intensity ratio for arc light spectra; and calculating short-time energy, zero-crossing rate, or envelope peak spacing for acoustic emission. The confidence level is determined by the normalized signal-to-noise ratio, cross-channel consistency measure, and missing rate.

3. The method according to claim 1, characterized in that, The start, pause, and return thresholds for sparse gating include an upper threshold and a lower threshold. The gating only enters activation when the weight rises above the upper threshold and exits activation when the weight falls below the lower threshold. The minimum dwell time is not less than 0.5 seconds and not more than 2 seconds.

4. The method according to claim 1, characterized in that, When constraints conflict, compliance agents prioritize and tailor the responses accordingly, with statutory standards taking precedence over corporate guidelines, which in turn take precedence over historical experience. They also adhere to the principle of minimal changes to satisfy all mandatory constraints.

5. The method according to claim 1, characterized in that, On-chain anchoring adopts a channel isolation strategy to distinguish different production lines or customers. Only summaries and necessary metadata are stored on-chain, while detailed data is stored off-chain in encrypted form and is subject to audit access control. Event types include at least parameter overruns, sensor mismatches, rework triggers, network anomalies, and billing settlements. Settlement is based on runtime and power consumption.

6. The method according to claim 1, characterized in that, The dual signature is completed separately in the compliance agent and the security agent using independent private keys. When the controller receives control instructions, it verifies the two signatures and timestamps and checks the clock deviation synchronized by the precision time protocol. If either verification fails, execution is blocked.

7. The method according to claim 1, characterized in that, The triggering conditions for degradation control include policy service timeout, model packet verification failure, and network quality indicators falling below a set threshold. Smooth switching uses time-varying weights to weight the degradation output and the gated output, with the weights transitioning from zero to one or from one to zero within the switching window.

8. A reliable multi-expert gated self-optimizing control system for a welding robot chain, characterized in that, include: a) Data acquisition and feature extraction module, used to acquire electrical participation motion signals and at least one molten pool related signal at a frequency of not less than 200 Hz, and output working condition characteristics and confidence level; b) A multi-expert control module is used to output candidate incremental suggestions and confidence levels for each control channel based on operating condition characteristics. The experts include at least molten pool stabilization, spatter suppression, heat input constraint and energy consumption optimization. c) A sparse gated routing module is used to select no more than two experts in each control cycle and linearly synthesize the comprehensive control increment according to weights, while performing hysteresis and minimum dwell constraints; d) A compliance agent module is used to implement unit unification, upper and lower bound pruning, maximum rate of change limit, and hot input interval constraint on the comprehensive control increment, and to match and verify the pose, floor, and bevel constraints, outputting the compliant control increment; e) A safety agent module is used to complete timing consistency, chattering, and actuator saturation judgment based on historical trajectory buffers, and to implement amplitude limiting, filtering, or freeze rollback for anomalies. The compliance agent and safety agent respectively process the final control increment. The system includes: (a) a digital signature module; (b) an execution and drive module, which writes the control increment to the controller after verifying the validity of the dual signature and implements hardware limiting and maximum slope constraints on the controller side; (c) a blockchain client module, which generates segmented summaries of control fragments and key quantities and performs Merkle aggregation on the chain, while generating and reporting events and linking with the settlement contract; (d) a degradation control module, which adopts a proportional-integral-derivative plus limiting and maximum slope strategy when the cloud-side policy is unavailable and performs time-weighted smooth switching after the cloud side recovers; and (e) a cloud-side policy service, which trains and issues expert and gating models, manages rule base versions, and provides an audit interface.

9. The system according to claim 8, characterized in that, Each expert in the multi-expert control module outputs the channel identifier, incremental value, and confidence level in a structured format. The sparse gating routing module assigns weights based on operating condition characteristics and confidence levels and limits the number of activated experts to no more than two.

10. The system according to claim 8, characterized in that, The rule base stores unit mappings, parameter upper and lower bounds, maximum rate of change, thermal input ranges for different material thicknesses and orientations, and matching rules for orientation and bevel. The execution and drive module configures hardware limiters and maximum rate of change constraints within the controller to form the final physical safety boundary.

Images